{ "compact_table": [ { "label": "cnnmca-table", "macro_success_lift_ci95_high": -0.043333333333333335, "macro_success_lift_ci95_low": -0.12179487179487179, "macro_success_lift_mean": -0.08358974358974358, "neg_hold_mean": 0.42059079285576884, "pos_drop_mean": 0.3959322889750625, "prob_delta_mean_mean": 0.010130963312934787, "t_cover_delta_mean": 0.0008494692000740305, "t_edge_delta_mean": -0.05989792490816818 }, { "label": "cnnmca-table-allgeom", "macro_success_lift_ci95_high": -0.041538461538461545, "macro_success_lift_ci95_low": -0.11358974358974357, "macro_success_lift_mean": -0.07794871794871795, "neg_hold_mean": 0.43696921600518773, "pos_drop_mean": 0.40278764503976827, "prob_delta_mean_mean": 0.011270323790640307, "t_cover_delta_mean": 0.009967706638735063, "t_edge_delta_mean": -0.05409223634163096 }, { "label": "dgcnn", "macro_success_lift_ci95_high": 0.10435897435897434, "macro_success_lift_ci95_low": 0.04769230769230768, "macro_success_lift_mean": 0.07538461538461538, "neg_hold_mean": 0.5316051405224628, "pos_drop_mean": 0.18741476354029776, "prob_delta_mean_mean": 0.06488625581080952, "t_cover_delta_mean": -0.018895160601424345, "t_edge_delta_mean": -0.05672891759318866 }, { "label": "full-sga-gsn", "macro_success_lift_ci95_high": 0.12461538461538461, "macro_success_lift_ci95_low": 0.07, "macro_success_lift_mean": 0.09666666666666666, "neg_hold_mean": 0.5612905806481303, "pos_drop_mean": 0.17209766636487392, "prob_delta_mean_mean": 0.04604703450734807, "t_cover_delta_mean": -0.0258658302826347, "t_edge_delta_mean": -0.05780233122614356 }, { "label": "full-sga-gsn-seed8", "macro_success_lift_ci95_high": 0.12282051282051283, "macro_success_lift_ci95_low": 0.07230769230769231, "macro_success_lift_mean": 0.09692307692307693, "neg_hold_mean": 0.5608627716522271, "pos_drop_mean": 0.16226287747165766, "prob_delta_mean_mean": 0.07111787822452087, "t_cover_delta_mean": -0.008115947057435615, "t_edge_delta_mean": -0.03709511615422506 }, { "label": "no-action", "macro_success_lift_ci95_high": 0.008461538461538463, "macro_success_lift_ci95_low": -0.010769230769230781, "macro_success_lift_mean": -0.0010256410256410246, "neg_hold_mean": 0.11824745529719692, "pos_drop_mean": 0.06768525244041983, "prob_delta_mean_mean": -0.012355186981927266, "t_cover_delta_mean": -0.047632351768154946, "t_edge_delta_mean": -0.09784604655701118 }, { "label": "rand-action", "macro_success_lift_ci95_high": -0.08692307692307694, "macro_success_lift_ci95_low": -0.13846153846153844, "macro_success_lift_mean": -0.11333333333333334, "neg_hold_mean": 0.25359166342128225, "pos_drop_mean": 0.3487118440140831, "prob_delta_mean_mean": -0.008880441190868659, "t_cover_delta_mean": -0.035898255547274924, "t_edge_delta_mean": -0.10212315514105379 }, { "label": "vanilla", "macro_success_lift_ci95_high": 0.1253846153846154, "macro_success_lift_ci95_low": 0.07615384615384614, "macro_success_lift_mean": 0.09999999999999999, "neg_hold_mean": 0.5483742849285207, "pos_drop_mean": 0.15261387402149754, "prob_delta_mean_mean": 0.14576396324894023, "t_cover_delta_mean": -0.029887289578148114, "t_edge_delta_mean": -0.062301181548525784 }, { "label": "wo-onl-cal", "macro_success_lift_ci95_high": 0.13102564102564104, "macro_success_lift_ci95_low": 0.07205128205128204, "macro_success_lift_mean": 0.10102564102564103, "neg_hold_mean": 0.5920216862507405, "pos_drop_mean": 0.1802449722027022, "prob_delta_mean_mean": 0.17591223537712355, "t_cover_delta_mean": -0.017471287380660495, "t_edge_delta_mean": -0.05709123373839061 }, { "label": "wo-stb-rwd", "macro_success_lift_ci95_high": 0.1292371794871794, "macro_success_lift_ci95_low": 0.07974358974358974, "macro_success_lift_mean": 0.10358974358974359, "neg_hold_mean": 0.5671671475135774, "pos_drop_mean": 0.15336418647453992, "prob_delta_mean_mean": 0.06732549386964481, "t_cover_delta_mean": -0.03074642589623998, "t_edge_delta_mean": -0.057859812806336536 }, { "label": "wo-tac-rwd", "macro_success_lift_ci95_high": 0.12666666666666668, "macro_success_lift_ci95_low": 0.07358974358974359, "macro_success_lift_mean": 0.09948717948717949, "neg_hold_mean": 0.5720496857008787, "pos_drop_mean": 0.17248415269306264, "prob_delta_mean_mean": 0.0741112461070111, "t_cover_delta_mean": -0.02270146052692559, "t_edge_delta_mean": -0.04943018935204376 }, { "label": "wo-tac-sem-rwd", "macro_success_lift_ci95_high": 0.11666666666666665, "macro_success_lift_ci95_low": 0.06230769230769229, "macro_success_lift_mean": 0.0887179487179487, "neg_hold_mean": 0.5682747745647158, "pos_drop_mean": 0.17934832438174986, "prob_delta_mean_mean": 0.08527541113834454, "t_cover_delta_mean": -0.02650838499759513, "t_edge_delta_mean": -0.061308455462895865 } ], "protocol": { "episodes_per_object": 100, "name": "unseen_test_formal_paper_seed7_episode", "primary_model": "full-sga-gsn-seed8", "seeds": [ 1001, 1002, 1003 ], "test_object_ids": [ 75, 76, 77, 79, 80, 81, 83, 84, 86 ] }, "rows": [ { "across_object_lift_iqr_mean": 0.19333333333333336, "across_object_lift_iqr_std": 0.07408703590297624, "across_object_lift_std_mean": 0.12430699569323393, "across_object_lift_std_std": 0.016777920978869976, "action_bin_rot_bin_0_count_mean": 0.0, "action_bin_rot_bin_0_count_std": 0.0, "action_bin_rot_bin_0_drop_rate_mean": null, "action_bin_rot_bin_0_drop_rate_std": null, "action_bin_rot_bin_0_prob_delta_mean_mean": null, "action_bin_rot_bin_0_prob_delta_mean_std": null, "action_bin_rot_bin_0_success_delta_mean_mean": null, "action_bin_rot_bin_0_success_delta_mean_std": null, "action_bin_rot_bin_1_count_mean": 0.0, "action_bin_rot_bin_1_count_std": 0.0, "action_bin_rot_bin_1_drop_rate_mean": null, "action_bin_rot_bin_1_drop_rate_std": null, "action_bin_rot_bin_1_prob_delta_mean_mean": null, "action_bin_rot_bin_1_prob_delta_mean_std": null, "action_bin_rot_bin_1_success_delta_mean_mean": null, "action_bin_rot_bin_1_success_delta_mean_std": null, "action_bin_rot_bin_2_count_mean": 0.0, "action_bin_rot_bin_2_count_std": 0.0, "action_bin_rot_bin_2_drop_rate_mean": null, "action_bin_rot_bin_2_drop_rate_std": null, "action_bin_rot_bin_2_prob_delta_mean_mean": null, "action_bin_rot_bin_2_prob_delta_mean_std": null, "action_bin_rot_bin_2_success_delta_mean_mean": null, "action_bin_rot_bin_2_success_delta_mean_std": null, "action_bin_rot_bin_3_count_mean": 1300.0, "action_bin_rot_bin_3_count_std": 0.0, "action_bin_rot_bin_3_drop_rate_mean": 0.2312820512820513, "action_bin_rot_bin_3_drop_rate_std": 0.005835798295764632, "action_bin_rot_bin_3_prob_delta_mean_mean": 0.010130963312934787, "action_bin_rot_bin_3_prob_delta_mean_std": 0.0006689933575558751, "action_bin_rot_bin_3_success_delta_mean_mean": -0.08358974358974358, "action_bin_rot_bin_3_success_delta_mean_std": 0.009448987575737576, "action_bin_trans_bin_0_count_mean": 0.0, "action_bin_trans_bin_0_count_std": 0.0, "action_bin_trans_bin_0_drop_rate_mean": null, "action_bin_trans_bin_0_drop_rate_std": null, "action_bin_trans_bin_0_prob_delta_mean_mean": null, "action_bin_trans_bin_0_prob_delta_mean_std": null, "action_bin_trans_bin_0_success_delta_mean_mean": null, "action_bin_trans_bin_0_success_delta_mean_std": null, "action_bin_trans_bin_1_count_mean": 0.0, "action_bin_trans_bin_1_count_std": 0.0, "action_bin_trans_bin_1_drop_rate_mean": null, "action_bin_trans_bin_1_drop_rate_std": null, "action_bin_trans_bin_1_prob_delta_mean_mean": null, "action_bin_trans_bin_1_prob_delta_mean_std": null, "action_bin_trans_bin_1_success_delta_mean_mean": null, "action_bin_trans_bin_1_success_delta_mean_std": null, "action_bin_trans_bin_2_count_mean": 0.0, "action_bin_trans_bin_2_count_std": 0.0, "action_bin_trans_bin_2_drop_rate_mean": null, "action_bin_trans_bin_2_drop_rate_std": null, "action_bin_trans_bin_2_prob_delta_mean_mean": null, "action_bin_trans_bin_2_prob_delta_mean_std": null, "action_bin_trans_bin_2_success_delta_mean_mean": null, "action_bin_trans_bin_2_success_delta_mean_std": null, "action_bin_trans_bin_3_count_mean": 1300.0, "action_bin_trans_bin_3_count_std": 0.0, "action_bin_trans_bin_3_drop_rate_mean": 0.2312820512820513, "action_bin_trans_bin_3_drop_rate_std": 0.005835798295764632, "action_bin_trans_bin_3_prob_delta_mean_mean": 0.010130963312934787, "action_bin_trans_bin_3_prob_delta_mean_std": 0.0006689933575558751, "action_bin_trans_bin_3_success_delta_mean_mean": -0.08358974358974358, "action_bin_trans_bin_3_success_delta_mean_std": 0.009448987575737576, "action_dim_0_mean_mean": 1.0, "action_dim_0_mean_std": 0.0, "action_dim_0_saturation_rate_mean": 1.0, "action_dim_0_saturation_rate_std": 0.0, "action_dim_0_std_mean": 0.0, "action_dim_0_std_std": 0.0, "action_dim_1_mean_mean": 1.0, "action_dim_1_mean_std": 0.0, "action_dim_1_saturation_rate_mean": 1.0, "action_dim_1_saturation_rate_std": 0.0, "action_dim_1_std_mean": 0.0, "action_dim_1_std_std": 0.0, "action_dim_2_mean_mean": 1.0, "action_dim_2_mean_std": 0.0, "action_dim_2_saturation_rate_mean": 1.0, "action_dim_2_saturation_rate_std": 0.0, "action_dim_2_std_mean": 0.0, "action_dim_2_std_std": 0.0, "action_dim_3_mean_mean": -1.0, "action_dim_3_mean_std": 0.0, "action_dim_3_saturation_rate_mean": 1.0, "action_dim_3_saturation_rate_std": 0.0, "action_dim_3_std_mean": 0.0, "action_dim_3_std_std": 0.0, "action_dim_4_mean_mean": -1.0, "action_dim_4_mean_std": 0.0, "action_dim_4_saturation_rate_mean": 1.0, "action_dim_4_saturation_rate_std": 0.0, "action_dim_4_std_mean": 0.0, "action_dim_4_std_std": 0.0, "action_dim_5_mean_mean": 1.0, "action_dim_5_mean_std": 0.0, "action_dim_5_saturation_rate_mean": 1.0, "action_dim_5_saturation_rate_std": 0.0, "action_dim_5_std_mean": 0.0, "action_dim_5_std_std": 0.0, "action_rotation_norm_mean_mean": 1.7320508075688774, "action_rotation_norm_mean_std": 2.220446049250313e-16, "action_rotation_norm_std_mean": 0.0, "action_rotation_norm_std_std": 0.0, "action_saturation_rate_mean": 1.0, "action_saturation_rate_std": 0.0, "action_translation_norm_mean_mean": 1.7320508075688774, "action_translation_norm_mean_std": 2.220446049250313e-16, "action_translation_norm_std_mean": 0.0, "action_translation_norm_std_std": 0.0, "calibrator_after_brier_mean": 0.24120486110615644, "calibrator_after_brier_std": 0.0004421213513783972, "calibrator_before_brier_vs_legacy_mean": 0.23097745165293584, "calibrator_before_brier_vs_legacy_std": 0.0010678455447441979, "calibrator_neg_prob_delta_degradation_auc_mean": 0.5714347289958054, "calibrator_neg_prob_delta_degradation_auc_std": 0.005593325860610293, "calibrator_prob_after_auc_mean": 0.647166657478229, "calibrator_prob_after_auc_std": 0.00751218294170213, "calibrator_prob_before_auc_vs_legacy_mean": 0.7936265972817765, "calibrator_prob_before_auc_vs_legacy_std": 0.00997076671981087, "calibrator_prob_delta_recovery_auc_mean": 0.5676596996920978, "calibrator_prob_delta_recovery_auc_std": 0.0076953223536891146, "calibrator_raw_logit_after_auc_mean": 0.647166657478229, "calibrator_raw_logit_after_auc_std": 0.00751218294170213, "calibrator_raw_logit_before_auc_vs_legacy_mean": 0.7936265972817765, "calibrator_raw_logit_before_auc_vs_legacy_std": 0.00997076671981087, "corr_dim_0_negative_recovery_mean": null, "corr_dim_0_negative_recovery_std": null, "corr_dim_0_positive_drop_mean": null, "corr_dim_0_positive_drop_std": null, "corr_dim_0_prob_delta_mean": null, "corr_dim_0_prob_delta_std": null, "corr_dim_1_negative_recovery_mean": null, "corr_dim_1_negative_recovery_std": null, "corr_dim_1_positive_drop_mean": null, "corr_dim_1_positive_drop_std": null, "corr_dim_1_prob_delta_mean": null, "corr_dim_1_prob_delta_std": null, "corr_dim_2_negative_recovery_mean": null, "corr_dim_2_negative_recovery_std": null, "corr_dim_2_positive_drop_mean": null, "corr_dim_2_positive_drop_std": null, "corr_dim_2_prob_delta_mean": null, "corr_dim_2_prob_delta_std": null, "corr_dim_3_negative_recovery_mean": null, "corr_dim_3_negative_recovery_std": null, "corr_dim_3_positive_drop_mean": null, "corr_dim_3_positive_drop_std": null, "corr_dim_3_prob_delta_mean": null, "corr_dim_3_prob_delta_std": null, "corr_dim_4_negative_recovery_mean": null, "corr_dim_4_negative_recovery_std": null, "corr_dim_4_positive_drop_mean": null, "corr_dim_4_positive_drop_std": null, "corr_dim_4_prob_delta_mean": null, "corr_dim_4_prob_delta_std": null, "corr_dim_5_negative_recovery_mean": null, "corr_dim_5_negative_recovery_std": null, "corr_dim_5_positive_drop_mean": null, "corr_dim_5_positive_drop_std": null, "corr_dim_5_prob_delta_mean": null, "corr_dim_5_prob_delta_std": null, "corr_rotation_norm_prob_delta_mean": null, "corr_rotation_norm_prob_delta_std": null, "corr_rotation_norm_success_delta_mean": null, "corr_rotation_norm_success_delta_std": null, "corr_translation_norm_prob_delta_mean": null, "corr_translation_norm_prob_delta_std": null, "corr_translation_norm_success_delta_mean": null, "corr_translation_norm_success_delta_std": null, "evaluation_wall_minutes": 30.04175890251839, "experiment_name": "cnnmca-table", "label": "cnnmca-table", "macro_success_lift_ci95_high": -0.043333333333333335, "macro_success_lift_ci95_low": -0.12179487179487179, "macro_success_lift_mean": -0.08358974358974358, "macro_success_lift_std": 0.00944898757573758, "neg_hold_ci95_high": 0.46128686880854, "neg_hold_ci95_low": 0.3807512802532878, "neg_hold_mean": 0.42059079285576884, "neg_hold_std": 0.029927589657718134, "pos_drop_ci95_high": 0.4215421424287043, "pos_drop_ci95_low": 0.37018368738555246, "pos_drop_mean": 0.3959322889750625, "pos_drop_std": 0.01321985915530301, "prob_delta_mean_ci95_high": 0.012107260629184453, "prob_delta_mean_ci95_low": 0.008181084512208454, "prob_delta_mean_mean": 0.010130963312934787, "prob_delta_mean_std": 0.0006689933575558759, "source_summary_csv": "/rl-grasp-refine/outputs/unseen_test_formal_paper_seed7_episode/cnnmca-table/summary.csv", "t_cover_delta_ci95_high": 0.01849754161720905, "t_cover_delta_ci95_low": -0.014164745920331059, "t_cover_delta_mean": 0.0008494692000740305, "t_cover_delta_std": 0.004283287885944958, "t_edge_delta_ci95_high": -0.039201014409108906, "t_edge_delta_ci95_low": -0.0773118311778787, "t_edge_delta_mean": -0.05989792490816818, "t_edge_delta_std": 0.008295669842051809 }, { "across_object_lift_iqr_mean": 0.18000000000000002, "across_object_lift_iqr_std": 0.056568542494923754, "across_object_lift_std_mean": 0.11731461137915689, "across_object_lift_std_std": 0.008646230515377615, "action_bin_rot_bin_0_count_mean": 0.0, "action_bin_rot_bin_0_count_std": 0.0, "action_bin_rot_bin_0_drop_rate_mean": null, "action_bin_rot_bin_0_drop_rate_std": null, "action_bin_rot_bin_0_prob_delta_mean_mean": null, "action_bin_rot_bin_0_prob_delta_mean_std": null, "action_bin_rot_bin_0_success_delta_mean_mean": null, "action_bin_rot_bin_0_success_delta_mean_std": null, "action_bin_rot_bin_1_count_mean": 0.0, "action_bin_rot_bin_1_count_std": 0.0, "action_bin_rot_bin_1_drop_rate_mean": null, "action_bin_rot_bin_1_drop_rate_std": null, "action_bin_rot_bin_1_prob_delta_mean_mean": null, "action_bin_rot_bin_1_prob_delta_mean_std": null, "action_bin_rot_bin_1_success_delta_mean_mean": null, "action_bin_rot_bin_1_success_delta_mean_std": null, "action_bin_rot_bin_2_count_mean": 0.0, "action_bin_rot_bin_2_count_std": 0.0, "action_bin_rot_bin_2_drop_rate_mean": null, "action_bin_rot_bin_2_drop_rate_std": null, "action_bin_rot_bin_2_prob_delta_mean_mean": null, "action_bin_rot_bin_2_prob_delta_mean_std": null, "action_bin_rot_bin_2_success_delta_mean_mean": null, "action_bin_rot_bin_2_success_delta_mean_std": null, "action_bin_rot_bin_3_count_mean": 1300.0, "action_bin_rot_bin_3_count_std": 0.0, "action_bin_rot_bin_3_drop_rate_mean": 0.22794871794871796, "action_bin_rot_bin_3_drop_rate_std": 0.005341709398974269, "action_bin_rot_bin_3_prob_delta_mean_mean": 0.011270323790640304, "action_bin_rot_bin_3_prob_delta_mean_std": 0.002471304296633012, "action_bin_rot_bin_3_success_delta_mean_mean": -0.07794871794871794, "action_bin_rot_bin_3_success_delta_mean_std": 0.010094881889758734, "action_bin_trans_bin_0_count_mean": 0.0, "action_bin_trans_bin_0_count_std": 0.0, "action_bin_trans_bin_0_drop_rate_mean": null, "action_bin_trans_bin_0_drop_rate_std": null, "action_bin_trans_bin_0_prob_delta_mean_mean": null, "action_bin_trans_bin_0_prob_delta_mean_std": null, "action_bin_trans_bin_0_success_delta_mean_mean": null, "action_bin_trans_bin_0_success_delta_mean_std": null, "action_bin_trans_bin_1_count_mean": 0.0, "action_bin_trans_bin_1_count_std": 0.0, "action_bin_trans_bin_1_drop_rate_mean": null, "action_bin_trans_bin_1_drop_rate_std": null, "action_bin_trans_bin_1_prob_delta_mean_mean": null, "action_bin_trans_bin_1_prob_delta_mean_std": null, "action_bin_trans_bin_1_success_delta_mean_mean": null, "action_bin_trans_bin_1_success_delta_mean_std": null, "action_bin_trans_bin_2_count_mean": 0.0, "action_bin_trans_bin_2_count_std": 0.0, "action_bin_trans_bin_2_drop_rate_mean": null, "action_bin_trans_bin_2_drop_rate_std": null, "action_bin_trans_bin_2_prob_delta_mean_mean": null, "action_bin_trans_bin_2_prob_delta_mean_std": null, "action_bin_trans_bin_2_success_delta_mean_mean": null, "action_bin_trans_bin_2_success_delta_mean_std": null, "action_bin_trans_bin_3_count_mean": 1300.0, "action_bin_trans_bin_3_count_std": 0.0, "action_bin_trans_bin_3_drop_rate_mean": 0.22794871794871796, "action_bin_trans_bin_3_drop_rate_std": 0.005341709398974269, "action_bin_trans_bin_3_prob_delta_mean_mean": 0.011270323790640304, "action_bin_trans_bin_3_prob_delta_mean_std": 0.002471304296633012, "action_bin_trans_bin_3_success_delta_mean_mean": -0.07794871794871794, "action_bin_trans_bin_3_success_delta_mean_std": 0.010094881889758734, "action_dim_0_mean_mean": 1.0, "action_dim_0_mean_std": 0.0, "action_dim_0_saturation_rate_mean": 1.0, "action_dim_0_saturation_rate_std": 0.0, "action_dim_0_std_mean": 0.0, "action_dim_0_std_std": 0.0, "action_dim_1_mean_mean": -1.0, "action_dim_1_mean_std": 0.0, "action_dim_1_saturation_rate_mean": 1.0, "action_dim_1_saturation_rate_std": 0.0, "action_dim_1_std_mean": 0.0, "action_dim_1_std_std": 0.0, "action_dim_2_mean_mean": -1.0, "action_dim_2_mean_std": 0.0, "action_dim_2_saturation_rate_mean": 1.0, "action_dim_2_saturation_rate_std": 0.0, "action_dim_2_std_mean": 0.0, "action_dim_2_std_std": 0.0, "action_dim_3_mean_mean": 1.0, "action_dim_3_mean_std": 0.0, "action_dim_3_saturation_rate_mean": 1.0, "action_dim_3_saturation_rate_std": 0.0, "action_dim_3_std_mean": 0.0, "action_dim_3_std_std": 0.0, "action_dim_4_mean_mean": 1.0, "action_dim_4_mean_std": 0.0, "action_dim_4_saturation_rate_mean": 1.0, "action_dim_4_saturation_rate_std": 0.0, "action_dim_4_std_mean": 0.0, "action_dim_4_std_std": 0.0, "action_dim_5_mean_mean": -1.0, "action_dim_5_mean_std": 0.0, "action_dim_5_saturation_rate_mean": 1.0, "action_dim_5_saturation_rate_std": 0.0, "action_dim_5_std_mean": 0.0, "action_dim_5_std_std": 0.0, "action_rotation_norm_mean_mean": 1.7320508075688774, "action_rotation_norm_mean_std": 2.220446049250313e-16, "action_rotation_norm_std_mean": 0.0, "action_rotation_norm_std_std": 0.0, "action_saturation_rate_mean": 1.0, "action_saturation_rate_std": 0.0, "action_translation_norm_mean_mean": 1.7320508075688774, "action_translation_norm_mean_std": 2.220446049250313e-16, "action_translation_norm_std_mean": 0.0, "action_translation_norm_std_std": 0.0, "calibrator_after_brier_mean": 0.23348667872153875, "calibrator_after_brier_std": 0.0008918426359832548, "calibrator_before_brier_vs_legacy_mean": 0.2139080562076643, "calibrator_before_brier_vs_legacy_std": 0.0015001988800254745, "calibrator_neg_prob_delta_degradation_auc_mean": 0.5903864514852176, "calibrator_neg_prob_delta_degradation_auc_std": 0.013432096270165698, "calibrator_prob_after_auc_mean": 0.6661848333666547, "calibrator_prob_after_auc_std": 0.003891051065351252, "calibrator_prob_before_auc_vs_legacy_mean": 0.792896560895001, "calibrator_prob_before_auc_vs_legacy_std": 0.007378758777483622, "calibrator_prob_delta_recovery_auc_mean": 0.531787107541361, "calibrator_prob_delta_recovery_auc_std": 0.03984107218437759, "calibrator_raw_logit_after_auc_mean": 0.6661848333666547, "calibrator_raw_logit_after_auc_std": 0.003891051065351252, "calibrator_raw_logit_before_auc_vs_legacy_mean": 0.792896560895001, "calibrator_raw_logit_before_auc_vs_legacy_std": 0.007378758777483622, "corr_dim_0_negative_recovery_mean": null, "corr_dim_0_negative_recovery_std": null, "corr_dim_0_positive_drop_mean": null, "corr_dim_0_positive_drop_std": null, "corr_dim_0_prob_delta_mean": null, "corr_dim_0_prob_delta_std": null, "corr_dim_1_negative_recovery_mean": null, "corr_dim_1_negative_recovery_std": null, "corr_dim_1_positive_drop_mean": null, "corr_dim_1_positive_drop_std": null, "corr_dim_1_prob_delta_mean": null, "corr_dim_1_prob_delta_std": null, "corr_dim_2_negative_recovery_mean": null, "corr_dim_2_negative_recovery_std": null, "corr_dim_2_positive_drop_mean": null, "corr_dim_2_positive_drop_std": null, "corr_dim_2_prob_delta_mean": null, "corr_dim_2_prob_delta_std": null, "corr_dim_3_negative_recovery_mean": null, "corr_dim_3_negative_recovery_std": null, "corr_dim_3_positive_drop_mean": null, "corr_dim_3_positive_drop_std": null, "corr_dim_3_prob_delta_mean": null, "corr_dim_3_prob_delta_std": null, "corr_dim_4_negative_recovery_mean": null, "corr_dim_4_negative_recovery_std": null, "corr_dim_4_positive_drop_mean": null, "corr_dim_4_positive_drop_std": null, "corr_dim_4_prob_delta_mean": null, "corr_dim_4_prob_delta_std": null, "corr_dim_5_negative_recovery_mean": null, "corr_dim_5_negative_recovery_std": null, "corr_dim_5_positive_drop_mean": null, "corr_dim_5_positive_drop_std": null, "corr_dim_5_prob_delta_mean": null, "corr_dim_5_prob_delta_std": null, "corr_rotation_norm_prob_delta_mean": null, "corr_rotation_norm_prob_delta_std": null, "corr_rotation_norm_success_delta_mean": null, "corr_rotation_norm_success_delta_std": null, "corr_translation_norm_prob_delta_mean": null, "corr_translation_norm_prob_delta_std": null, "corr_translation_norm_success_delta_mean": null, "corr_translation_norm_success_delta_std": null, "evaluation_wall_minutes": 36.882084511212696, "experiment_name": "cnnmca-table-allgeom", "label": "cnnmca-table-allgeom", "macro_success_lift_ci95_high": -0.041538461538461545, "macro_success_lift_ci95_low": -0.11358974358974357, "macro_success_lift_mean": -0.07794871794871795, "macro_success_lift_std": 0.010094881889758716, "neg_hold_ci95_high": 0.4824296554180737, "neg_hold_ci95_low": 0.38997607408421087, "neg_hold_mean": 0.43696921600518773, "neg_hold_std": 0.018538715301322135, "pos_drop_ci95_high": 0.441461740075072, "pos_drop_ci95_low": 0.3661836557319038, "pos_drop_mean": 0.40278764503976827, "pos_drop_std": 0.01937826156431396, "prob_delta_mean_ci95_high": 0.014483455517812119, "prob_delta_mean_ci95_low": 0.008077139087236361, "prob_delta_mean_mean": 0.011270323790640307, "prob_delta_mean_std": 0.002471304296633013, "source_summary_csv": "/rl-grasp-refine/outputs/unseen_test_formal_paper_seed7_episode/cnnmca-table-allgeom/summary.csv", "t_cover_delta_ci95_high": 0.02406928107330437, "t_cover_delta_ci95_low": -0.002861123394270364, "t_cover_delta_mean": 0.009967706638735063, "t_cover_delta_std": 0.00777680559575292, "t_edge_delta_ci95_high": -0.03629154044076563, "t_edge_delta_ci95_low": -0.07025146305288907, "t_edge_delta_mean": -0.05409223634163096, "t_edge_delta_std": 0.005813016001101493 }, { "across_object_lift_iqr_mean": 0.12666666666666668, "across_object_lift_iqr_std": 0.041096093353126514, "across_object_lift_std_mean": 0.08940317067060516, "across_object_lift_std_std": 0.016341163554376658, "action_bin_rot_bin_0_count_mean": 632.0, "action_bin_rot_bin_0_count_std": 9.201449161228174, "action_bin_rot_bin_0_drop_rate_mean": 0.12708293047848065, "action_bin_rot_bin_0_drop_rate_std": 0.0023380961790815612, "action_bin_rot_bin_0_prob_delta_mean_mean": 0.004115153548366998, "action_bin_rot_bin_0_prob_delta_mean_std": 0.0031229631235324203, "action_bin_rot_bin_0_success_delta_mean_mean": -0.035795096898052436, "action_bin_rot_bin_0_success_delta_mean_std": 0.004911595092117765, "action_bin_rot_bin_1_count_mean": 236.66666666666666, "action_bin_rot_bin_1_count_std": 11.469767022723502, "action_bin_rot_bin_1_drop_rate_mean": 0.11479254043959926, "action_bin_rot_bin_1_drop_rate_std": 0.03076890871989714, "action_bin_rot_bin_1_prob_delta_mean_mean": 0.06269304873760645, "action_bin_rot_bin_1_prob_delta_mean_std": 0.0023866140477726257, "action_bin_rot_bin_1_success_delta_mean_mean": 0.09427703333585687, "action_bin_rot_bin_1_success_delta_mean_std": 0.034299038143391365, "action_bin_rot_bin_2_count_mean": 162.66666666666666, "action_bin_rot_bin_2_count_std": 3.39934634239519, "action_bin_rot_bin_2_drop_rate_mean": 0.10488387635134548, "action_bin_rot_bin_2_drop_rate_std": 0.02000464295386719, "action_bin_rot_bin_2_prob_delta_mean_mean": 0.11199161259713987, "action_bin_rot_bin_2_prob_delta_mean_std": 0.005175927212077661, "action_bin_rot_bin_2_success_delta_mean_mean": 0.21096984418067455, "action_bin_rot_bin_2_success_delta_mean_std": 0.03127416602678905, "action_bin_rot_bin_3_count_mean": 268.6666666666667, "action_bin_rot_bin_3_count_std": 3.0912061651652345, "action_bin_rot_bin_3_drop_rate_mean": 0.047089417580251404, "action_bin_rot_bin_3_drop_rate_std": 0.015535455979987112, "action_bin_rot_bin_3_prob_delta_mean_mean": 0.18119245769854167, "action_bin_rot_bin_3_prob_delta_mean_std": 0.0068954426448030615, "action_bin_rot_bin_3_success_delta_mean_mean": 0.23843424464358823, "action_bin_rot_bin_3_success_delta_mean_std": 0.04817706494344496, "action_bin_trans_bin_0_count_mean": 0.0, "action_bin_trans_bin_0_count_std": 0.0, "action_bin_trans_bin_0_drop_rate_mean": null, "action_bin_trans_bin_0_drop_rate_std": null, "action_bin_trans_bin_0_prob_delta_mean_mean": null, "action_bin_trans_bin_0_prob_delta_mean_std": null, "action_bin_trans_bin_0_success_delta_mean_mean": null, "action_bin_trans_bin_0_success_delta_mean_std": null, "action_bin_trans_bin_1_count_mean": 508.0, "action_bin_trans_bin_1_count_std": 11.575836902790225, "action_bin_trans_bin_1_drop_rate_mean": 0.11816489493022571, "action_bin_trans_bin_1_drop_rate_std": 0.002868774322690758, "action_bin_trans_bin_1_prob_delta_mean_mean": -0.0036784251809178583, "action_bin_trans_bin_1_prob_delta_mean_std": 0.0035334438420984446, "action_bin_trans_bin_1_success_delta_mean_mean": -0.043914361751196195, "action_bin_trans_bin_1_success_delta_mean_std": 0.006179698213417934, "action_bin_trans_bin_2_count_mean": 445.6666666666667, "action_bin_trans_bin_2_count_std": 10.208928554075703, "action_bin_trans_bin_2_drop_rate_mean": 0.11401775188960543, "action_bin_trans_bin_2_drop_rate_std": 0.020465401526396122, "action_bin_trans_bin_2_prob_delta_mean_mean": 0.06960399198697619, "action_bin_trans_bin_2_prob_delta_mean_std": 0.0040151055404902105, "action_bin_trans_bin_2_success_delta_mean_mean": 0.10399244157825394, "action_bin_trans_bin_2_success_delta_mean_std": 0.012395647507397058, "action_bin_trans_bin_3_count_mean": 346.3333333333333, "action_bin_trans_bin_3_count_std": 1.699673171197595, "action_bin_trans_bin_3_drop_rate_mean": 0.07601795381602365, "action_bin_trans_bin_3_drop_rate_std": 0.00517911036062396, "action_bin_trans_bin_3_prob_delta_mean_mean": 0.1595530597725935, "action_bin_trans_bin_3_prob_delta_mean_std": 0.005498302025495972, "action_bin_trans_bin_3_success_delta_mean_mean": 0.21370536806747978, "action_bin_trans_bin_3_success_delta_mean_std": 0.020449599472363973, "action_dim_0_mean_mean": 0.8185113009657615, "action_dim_0_mean_std": 0.002090320490940424, "action_dim_0_saturation_rate_mean": 0.5451282051282051, "action_dim_0_saturation_rate_std": 0.005835798295764623, "action_dim_0_std_mean": 0.20387331192747413, "action_dim_0_std_std": 0.000531593898469525, "action_dim_1_mean_mean": 0.18311415316631052, "action_dim_1_mean_std": 0.0030016332418898294, "action_dim_1_saturation_rate_mean": 0.09076923076923077, "action_dim_1_saturation_rate_std": 0.006527139518645058, "action_dim_1_std_mean": 0.37519387752238803, "action_dim_1_std_std": 0.0028530875870414036, "action_dim_2_mean_mean": 0.20465029453004271, "action_dim_2_mean_std": 0.011193421353548897, "action_dim_2_saturation_rate_mean": 0.1912820512820513, "action_dim_2_saturation_rate_std": 0.012907397896054586, "action_dim_2_std_mean": 0.5088320072226521, "action_dim_2_std_std": 0.004020097313029775, "action_dim_3_mean_mean": 0.2829703000678236, "action_dim_3_mean_std": 0.0034944168339368156, "action_dim_3_saturation_rate_mean": 0.2271794871794872, "action_dim_3_saturation_rate_std": 0.0020189763779517537, "action_dim_3_std_mean": 0.4886439216358586, "action_dim_3_std_std": 4.4575582010070075e-05, "action_dim_4_mean_mean": -0.13325625411982542, "action_dim_4_mean_std": 0.006243146769258815, "action_dim_4_saturation_rate_mean": 0.12717948717948718, "action_dim_4_saturation_rate_std": 0.005699002761767404, "action_dim_4_std_mean": 0.4830338998458736, "action_dim_4_std_std": 0.006539680334600616, "action_dim_5_mean_mean": 0.16511412443801346, "action_dim_5_mean_std": 0.0012350857023457644, "action_dim_5_saturation_rate_mean": 0.11025641025641025, "action_dim_5_saturation_rate_std": 0.005629358564128245, "action_dim_5_std_mean": 0.4183976282690192, "action_dim_5_std_std": 0.0024824673837445135, "action_rotation_norm_mean_mean": 0.7024777480762925, "action_rotation_norm_mean_std": 0.00405751513958853, "action_rotation_norm_std_mean": 0.5280533049670479, "action_rotation_norm_std_std": 0.0036335139969215923, "action_saturation_rate_mean": 0.2152991452991453, "action_saturation_rate_std": 0.004512545316654457, "action_translation_norm_mean_mean": 1.0140345795738088, "action_translation_norm_mean_std": 0.003961297327183857, "action_translation_norm_std_mean": 0.39811594604272393, "action_translation_norm_std_std": 0.001788696675923867, "calibrator_after_brier_mean": 0.19174975483610188, "calibrator_after_brier_std": 0.0010312055007291807, "calibrator_before_brier_vs_legacy_mean": 0.17542041535533381, "calibrator_before_brier_vs_legacy_std": 0.0005251004426055665, "calibrator_neg_prob_delta_degradation_auc_mean": 0.5560696490481117, "calibrator_neg_prob_delta_degradation_auc_std": 0.018342901023793663, "calibrator_prob_after_auc_mean": 0.747286163807349, "calibrator_prob_after_auc_std": 0.005463540071660345, "calibrator_prob_before_auc_vs_legacy_mean": 0.8607725786172988, "calibrator_prob_before_auc_vs_legacy_std": 0.002413835973086723, "calibrator_prob_delta_recovery_auc_mean": 0.5070553193926813, "calibrator_prob_delta_recovery_auc_std": 0.028460321179396217, "calibrator_raw_logit_after_auc_mean": 0.747286163807349, "calibrator_raw_logit_after_auc_std": 0.005463540071660345, "calibrator_raw_logit_before_auc_vs_legacy_mean": 0.8607725786172988, "calibrator_raw_logit_before_auc_vs_legacy_std": 0.002413835973086723, "corr_dim_0_negative_recovery_mean": -0.06557878003037541, "corr_dim_0_negative_recovery_std": 0.042538824657806826, "corr_dim_0_positive_drop_mean": 0.12077125890044221, "corr_dim_0_positive_drop_std": 0.02660749394492738, "corr_dim_0_prob_delta_mean": 0.5222800145408393, "corr_dim_0_prob_delta_std": 0.015561583246972224, "corr_dim_1_negative_recovery_mean": -0.22305942547730315, "corr_dim_1_negative_recovery_std": 0.012774571346603224, "corr_dim_1_positive_drop_mean": 0.14366818939982592, "corr_dim_1_positive_drop_std": 0.014890571637051288, "corr_dim_1_prob_delta_mean": 0.5815641863997896, "corr_dim_1_prob_delta_std": 0.017733216968759787, "corr_dim_2_negative_recovery_mean": -0.1534888497198589, "corr_dim_2_negative_recovery_std": 0.02031356920745465, "corr_dim_2_positive_drop_mean": 0.10400600787208582, "corr_dim_2_positive_drop_std": 0.025765105059837524, "corr_dim_2_prob_delta_mean": 0.3239237680275202, "corr_dim_2_prob_delta_std": 0.01473165497666606, "corr_dim_3_negative_recovery_mean": -0.16158886654248286, "corr_dim_3_negative_recovery_std": 0.028168709621325068, "corr_dim_3_positive_drop_mean": 0.12556443731747688, "corr_dim_3_positive_drop_std": 0.005459314206240017, "corr_dim_3_prob_delta_mean": 0.6049308861425202, "corr_dim_3_prob_delta_std": 0.013518389146399844, "corr_dim_4_negative_recovery_mean": 0.2186971467395242, "corr_dim_4_negative_recovery_std": 0.013077755447067827, "corr_dim_4_positive_drop_mean": -0.06471416363595535, "corr_dim_4_positive_drop_std": 0.04449916172074677, "corr_dim_4_prob_delta_mean": -0.4038056543243467, "corr_dim_4_prob_delta_std": 0.0129388955213492, "corr_dim_5_negative_recovery_mean": -0.22620195215369052, "corr_dim_5_negative_recovery_std": 0.025225231172924602, "corr_dim_5_positive_drop_mean": 0.10339103775648006, "corr_dim_5_positive_drop_std": 0.04171873239456776, "corr_dim_5_prob_delta_mean": 0.5731417319358704, "corr_dim_5_prob_delta_std": 0.012240009364278262, "corr_rotation_norm_prob_delta_mean": 0.6180528713085834, "corr_rotation_norm_prob_delta_std": 0.01752245756611696, "corr_rotation_norm_success_delta_mean": 0.20981098800624845, "corr_rotation_norm_success_delta_std": 0.031725955914656795, "corr_translation_norm_prob_delta_mean": 0.5995622291609822, "corr_translation_norm_prob_delta_std": 0.017102869249593367, "corr_translation_norm_success_delta_mean": 0.2017096335634301, "corr_translation_norm_success_delta_std": 0.012378898840825395, "evaluation_wall_minutes": 36.85270333648659, "experiment_name": "dgcnn", "label": "dgcnn", "macro_success_lift_ci95_high": 0.10435897435897434, "macro_success_lift_ci95_low": 0.04769230769230768, "macro_success_lift_mean": 0.07538461538461538, "macro_success_lift_std": 0.0033234567684142847, "neg_hold_ci95_high": 0.5921522412659203, "neg_hold_ci95_low": 0.4728002460940859, "neg_hold_mean": 0.5316051405224628, "neg_hold_std": 0.02952560173274154, "pos_drop_ci95_high": 0.21675919025740314, "pos_drop_ci95_low": 0.15928001873882885, "pos_drop_mean": 0.18741476354029776, "pos_drop_std": 0.007470128591490997, "prob_delta_mean_ci95_high": 0.07396862101136747, "prob_delta_mean_ci95_low": 0.05679621710264345, "prob_delta_mean_mean": 0.06488625581080952, "prob_delta_mean_std": 0.0014568752862099527, "source_summary_csv": "/rl-grasp-refine/outputs/unseen_test_formal_paper_seed7_episode/dgcnn/summary.csv", "t_cover_delta_ci95_high": -0.006035664879768098, "t_cover_delta_ci95_low": -0.02928440834417002, "t_cover_delta_mean": -0.018895160601424345, "t_cover_delta_std": 0.005486551125036692, "t_edge_delta_ci95_high": -0.038508898919468966, "t_edge_delta_ci95_low": -0.07265668466726119, "t_edge_delta_mean": -0.05672891759318866, "t_edge_delta_std": 0.008272482460318116 }, { "across_object_lift_iqr_mean": 0.12666666666666668, "across_object_lift_iqr_std": 0.04496912521077342, "across_object_lift_std_mean": 0.08741998805520457, "across_object_lift_std_std": 0.007185491632852823, "action_bin_rot_bin_0_count_mean": 93.33333333333333, "action_bin_rot_bin_0_count_std": 9.672412085697939, "action_bin_rot_bin_0_drop_rate_mean": 0.08969647611172317, "action_bin_rot_bin_0_drop_rate_std": 0.0039727691613915955, "action_bin_rot_bin_0_prob_delta_mean_mean": 0.009205306346437223, "action_bin_rot_bin_0_prob_delta_mean_std": 0.0015968669184607774, "action_bin_rot_bin_0_success_delta_mean_mean": -0.021439200706826183, "action_bin_rot_bin_0_success_delta_mean_std": 0.03618009161853571, "action_bin_rot_bin_1_count_mean": 702.6666666666666, "action_bin_rot_bin_1_count_std": 6.128258770283412, "action_bin_rot_bin_1_drop_rate_mean": 0.1110710477958266, "action_bin_rot_bin_1_drop_rate_std": 0.008004858001329017, "action_bin_rot_bin_1_prob_delta_mean_mean": 0.035173743344364156, "action_bin_rot_bin_1_prob_delta_mean_std": 0.001307728914427135, "action_bin_rot_bin_1_success_delta_mean_mean": 0.022664563438704693, "action_bin_rot_bin_1_success_delta_mean_std": 0.012414653011490227, "action_bin_rot_bin_2_count_mean": 504.0, "action_bin_rot_bin_2_count_std": 11.430952132988164, "action_bin_rot_bin_2_drop_rate_mean": 0.08284406855835429, "action_bin_rot_bin_2_drop_rate_std": 0.007977290049371432, "action_bin_rot_bin_2_prob_delta_mean_mean": 0.06809441132297867, "action_bin_rot_bin_2_prob_delta_mean_std": 0.0039326858097466395, "action_bin_rot_bin_2_success_delta_mean_mean": 0.22088466017037445, "action_bin_rot_bin_2_success_delta_mean_std": 0.0007971539798348018, "action_bin_rot_bin_3_count_mean": 0.0, "action_bin_rot_bin_3_count_std": 0.0, "action_bin_rot_bin_3_drop_rate_mean": null, "action_bin_rot_bin_3_drop_rate_std": null, "action_bin_rot_bin_3_prob_delta_mean_mean": null, "action_bin_rot_bin_3_prob_delta_mean_std": null, "action_bin_rot_bin_3_success_delta_mean_mean": null, "action_bin_rot_bin_3_success_delta_mean_std": null, "action_bin_trans_bin_0_count_mean": 60.666666666666664, "action_bin_trans_bin_0_count_std": 6.182412330330469, "action_bin_trans_bin_0_drop_rate_mean": 0.09111548174048174, "action_bin_trans_bin_0_drop_rate_std": 0.0315796292142748, "action_bin_trans_bin_0_prob_delta_mean_mean": -0.024426265004046138, "action_bin_trans_bin_0_prob_delta_mean_std": 0.0016304002384624756, "action_bin_trans_bin_0_success_delta_mean_mean": -0.07564831002331002, "action_bin_trans_bin_0_success_delta_mean_std": 0.041700050657358455, "action_bin_trans_bin_1_count_mean": 282.0, "action_bin_trans_bin_1_count_std": 16.268579122549905, "action_bin_trans_bin_1_drop_rate_mean": 0.12407299118528525, "action_bin_trans_bin_1_drop_rate_std": 0.013401247815818619, "action_bin_trans_bin_1_prob_delta_mean_mean": -0.0004871193414948389, "action_bin_trans_bin_1_prob_delta_mean_std": 0.0024317835232650874, "action_bin_trans_bin_1_success_delta_mean_mean": -0.06675060921829619, "action_bin_trans_bin_1_success_delta_mean_std": 0.020642182869278394, "action_bin_trans_bin_2_count_mean": 741.3333333333334, "action_bin_trans_bin_2_count_std": 17.594190960528863, "action_bin_trans_bin_2_drop_rate_mean": 0.1002719245443499, "action_bin_trans_bin_2_drop_rate_std": 0.00011644938217040959, "action_bin_trans_bin_2_prob_delta_mean_mean": 0.05474804174407636, "action_bin_trans_bin_2_prob_delta_mean_std": 0.0008788592538617293, "action_bin_trans_bin_2_success_delta_mean_mean": 0.12871603364353848, "action_bin_trans_bin_2_success_delta_mean_std": 0.005156161352454052, "action_bin_trans_bin_3_count_mean": 216.0, "action_bin_trans_bin_3_count_std": 2.8284271247461903, "action_bin_trans_bin_3_drop_rate_mean": 0.061880487114131966, "action_bin_trans_bin_3_drop_rate_std": 0.011614894734868236, "action_bin_trans_bin_3_prob_delta_mean_mean": 0.09661222882147047, "action_bin_trans_bin_3_prob_delta_mean_std": 0.002878718522639521, "action_bin_trans_bin_3_success_delta_mean_mean": 0.24662984990087797, "action_bin_trans_bin_3_success_delta_mean_std": 0.023674377873027435, "action_dim_0_mean_mean": 0.7744583214957936, "action_dim_0_mean_std": 0.006633067416026908, "action_dim_0_saturation_rate_mean": 0.40487179487179487, "action_dim_0_saturation_rate_std": 0.008911870357427049, "action_dim_0_std_mean": 0.2162448539158454, "action_dim_0_std_std": 0.007092197883482686, "action_dim_1_mean_mean": 0.13272576578498, "action_dim_1_mean_std": 0.0059359384233888205, "action_dim_1_saturation_rate_mean": 0.0041025641025641026, "action_dim_1_saturation_rate_std": 0.0003626188621469475, "action_dim_1_std_mean": 0.30638126221961487, "action_dim_1_std_std": 0.002394995028890482, "action_dim_2_mean_mean": -0.05718311867103554, "action_dim_2_mean_std": 0.0016197860372132852, "action_dim_2_saturation_rate_mean": 0.09717948717948717, "action_dim_2_saturation_rate_std": 0.0063224759004810015, "action_dim_2_std_mean": 0.5916868518154305, "action_dim_2_std_std": 0.004741994083334868, "action_dim_3_mean_mean": 0.15920819732581434, "action_dim_3_mean_std": 0.00946504023784459, "action_dim_3_saturation_rate_mean": 0.005384615384615384, "action_dim_3_saturation_rate_std": 0.003323456768414287, "action_dim_3_std_mean": 0.402791164601815, "action_dim_3_std_std": 0.0069933683575748775, "action_dim_4_mean_mean": 0.2117427875119924, "action_dim_4_mean_std": 0.007822793270432938, "action_dim_4_saturation_rate_mean": 0.0, "action_dim_4_saturation_rate_std": 0.0, "action_dim_4_std_mean": 0.3681286825351908, "action_dim_4_std_std": 0.00711900259604379, "action_dim_5_mean_mean": 0.47716132086015134, "action_dim_5_mean_std": 0.0047883772206583795, "action_dim_5_saturation_rate_mean": 0.0005128205128205128, "action_dim_5_saturation_rate_std": 0.0007252377242938949, "action_dim_5_std_mean": 0.21157395791067804, "action_dim_5_std_std": 0.0037504317803498076, "action_rotation_norm_mean_mean": 0.7725440068654003, "action_rotation_norm_mean_std": 0.005143814505601503, "action_rotation_norm_std_mean": 0.20930654939310198, "action_rotation_norm_std_std": 0.003950306500324134, "action_saturation_rate_mean": 0.08534188034188035, "action_saturation_rate_std": 0.0019367981527563536, "action_translation_norm_mean_mean": 1.0139859612185391, "action_translation_norm_mean_std": 0.008046754107843557, "action_translation_norm_std_mean": 0.2885895312953054, "action_translation_norm_std_std": 0.0063683636574220605, "calibrator_after_brier_mean": 0.19929205531696645, "calibrator_after_brier_std": 0.0020146262330847497, "calibrator_before_brier_vs_legacy_mean": 0.1933413109176155, "calibrator_before_brier_vs_legacy_std": 0.0006424262316956972, "calibrator_neg_prob_delta_degradation_auc_mean": 0.5800612204237745, "calibrator_neg_prob_delta_degradation_auc_std": 0.014877810696290154, "calibrator_prob_after_auc_mean": 0.7309479024532979, "calibrator_prob_after_auc_std": 0.020691615213393983, "calibrator_prob_before_auc_vs_legacy_mean": 0.8625563400907638, "calibrator_prob_before_auc_vs_legacy_std": 0.004937261345008804, "calibrator_prob_delta_recovery_auc_mean": 0.5193193699630294, "calibrator_prob_delta_recovery_auc_std": 0.013705655065468248, "calibrator_raw_logit_after_auc_mean": 0.7309479024532979, "calibrator_raw_logit_after_auc_std": 0.020691615213393983, "calibrator_raw_logit_before_auc_vs_legacy_mean": 0.8625563400907638, "calibrator_raw_logit_before_auc_vs_legacy_std": 0.004937261345008804, "corr_dim_0_negative_recovery_mean": -0.1725773125050015, "corr_dim_0_negative_recovery_std": 0.039559313203809524, "corr_dim_0_positive_drop_mean": 0.07661989514052493, "corr_dim_0_positive_drop_std": 0.03906606478310347, "corr_dim_0_prob_delta_mean": 0.44846074982480416, "corr_dim_0_prob_delta_std": 0.01408390838049499, "corr_dim_1_negative_recovery_mean": 0.14253786325104922, "corr_dim_1_negative_recovery_std": 0.021789016179439038, "corr_dim_1_positive_drop_mean": 0.023512772938741516, "corr_dim_1_positive_drop_std": 0.019496629187428238, "corr_dim_1_prob_delta_mean": 0.055925882243462115, "corr_dim_1_prob_delta_std": 0.004563505636554912, "corr_dim_2_negative_recovery_mean": -0.13882016827118868, "corr_dim_2_negative_recovery_std": 0.031088354682580997, "corr_dim_2_positive_drop_mean": -0.005343312635600766, "corr_dim_2_positive_drop_std": 0.037110846564403145, "corr_dim_2_prob_delta_mean": 0.16675052921244085, "corr_dim_2_prob_delta_std": 0.0167882190125802, "corr_dim_3_negative_recovery_mean": 0.11823700897124108, "corr_dim_3_negative_recovery_std": 0.020202939100306214, "corr_dim_3_positive_drop_mean": 0.021767976101681658, "corr_dim_3_positive_drop_std": 0.03180123643338696, "corr_dim_3_prob_delta_mean": 0.275721679044246, "corr_dim_3_prob_delta_std": 0.008602933617166311, "corr_dim_4_negative_recovery_mean": 0.028950875895425443, "corr_dim_4_negative_recovery_std": 0.03725803431152237, "corr_dim_4_positive_drop_mean": 0.036895939148348565, "corr_dim_4_positive_drop_std": 0.027501192783756335, "corr_dim_4_prob_delta_mean": 0.3212510533914324, "corr_dim_4_prob_delta_std": 0.017912520489789665, "corr_dim_5_negative_recovery_mean": 0.03283899279973169, "corr_dim_5_negative_recovery_std": 0.019565478350828655, "corr_dim_5_positive_drop_mean": -0.002861274741784723, "corr_dim_5_positive_drop_std": 0.04837313326124955, "corr_dim_5_prob_delta_mean": 0.34843027054704345, "corr_dim_5_prob_delta_std": 0.011104373192434786, "corr_rotation_norm_prob_delta_mean": 0.2815646806099359, "corr_rotation_norm_prob_delta_std": 0.021550006917817083, "corr_rotation_norm_success_delta_mean": 0.19589031839547885, "corr_rotation_norm_success_delta_std": 0.009014551144158528, "corr_translation_norm_prob_delta_mean": 0.4685211637572708, "corr_translation_norm_prob_delta_std": 0.00511039332195132, "corr_translation_norm_success_delta_mean": 0.2088081539699627, "corr_translation_norm_success_delta_std": 0.010667306986806827, "evaluation_wall_minutes": 47.03731832893876, "experiment_name": "full-sga-gsn", "label": "full-sga-gsn", "macro_success_lift_ci95_high": 0.12461538461538461, "macro_success_lift_ci95_low": 0.07, "macro_success_lift_mean": 0.09666666666666666, "macro_success_lift_std": 0.006415382565741955, "neg_hold_ci95_high": 0.6119858801455378, "neg_hold_ci95_low": 0.5101995959522506, "neg_hold_mean": 0.5612905806481303, "neg_hold_std": 0.0070891947761534445, "pos_drop_ci95_high": 0.19785534675629005, "pos_drop_ci95_low": 0.1485031254550907, "pos_drop_mean": 0.17209766636487392, "pos_drop_std": 0.00778723784795693, "prob_delta_mean_ci95_high": 0.05032883354484278, "prob_delta_mean_ci95_low": 0.04207423363961618, "prob_delta_mean_mean": 0.04604703450734807, "prob_delta_mean_std": 0.0004119955059163138, "source_summary_csv": "/rl-grasp-refine/outputs/unseen_test_formal_paper_seed7_episode/full-sga-gsn/summary.csv", "t_cover_delta_ci95_high": -0.017191084543849013, "t_cover_delta_ci95_low": -0.03324666093443006, "t_cover_delta_mean": -0.0258658302826347, "t_cover_delta_std": 0.00044145246057812875, "t_edge_delta_ci95_high": -0.04382147330461139, "t_edge_delta_ci95_low": -0.07136968883302418, "t_edge_delta_mean": -0.05780233122614356, "t_edge_delta_std": 0.0049540362393037055 }, { "across_object_lift_iqr_mean": 0.1166666666666667, "across_object_lift_iqr_std": 0.016996731711976014, "across_object_lift_std_mean": 0.07919524028201462, "across_object_lift_std_std": 0.01342329775394745, "action_bin_rot_bin_0_count_mean": 284.0, "action_bin_rot_bin_0_count_std": 6.164414002968976, "action_bin_rot_bin_0_drop_rate_mean": 0.09892542876689707, "action_bin_rot_bin_0_drop_rate_std": 0.015429270653557455, "action_bin_rot_bin_0_prob_delta_mean_mean": 0.01201623560193775, "action_bin_rot_bin_0_prob_delta_mean_std": 0.0049345839584556, "action_bin_rot_bin_0_success_delta_mean_mean": -0.009892896742079313, "action_bin_rot_bin_0_success_delta_mean_std": 0.02663806125217466, "action_bin_rot_bin_1_count_mean": 644.0, "action_bin_rot_bin_1_count_std": 11.775681155103795, "action_bin_rot_bin_1_drop_rate_mean": 0.12525112677406983, "action_bin_rot_bin_1_drop_rate_std": 0.007899264339652714, "action_bin_rot_bin_1_prob_delta_mean_mean": 0.047332112174386286, "action_bin_rot_bin_1_prob_delta_mean_std": 0.000618354071544978, "action_bin_rot_bin_1_success_delta_mean_mean": 0.041434319460427055, "action_bin_rot_bin_1_success_delta_mean_std": 0.001459145225009157, "action_bin_rot_bin_2_count_mean": 232.66666666666666, "action_bin_rot_bin_2_count_std": 5.436502143433363, "action_bin_rot_bin_2_drop_rate_mean": 0.05129856095099605, "action_bin_rot_bin_2_drop_rate_std": 0.015135666605747033, "action_bin_rot_bin_2_prob_delta_mean_mean": 0.12499967973492887, "action_bin_rot_bin_2_prob_delta_mean_std": 0.009101057162342134, "action_bin_rot_bin_2_success_delta_mean_mean": 0.2843288278624043, "action_bin_rot_bin_2_success_delta_mean_std": 0.048896898319929195, "action_bin_rot_bin_3_count_mean": 139.33333333333334, "action_bin_rot_bin_3_count_std": 9.843215373488933, "action_bin_rot_bin_3_drop_rate_mean": 0.009183114035087718, "action_bin_rot_bin_3_drop_rate_std": 0.00811560928897339, "action_bin_rot_bin_3_prob_delta_mean_mean": 0.21172275379329897, "action_bin_rot_bin_3_prob_delta_mean_std": 0.0038860324005540567, "action_bin_rot_bin_3_success_delta_mean_mean": 0.2587302154843631, "action_bin_rot_bin_3_success_delta_mean_std": 0.03027704789365456, "action_bin_trans_bin_0_count_mean": 22.0, "action_bin_trans_bin_0_count_std": 1.632993161855452, "action_bin_trans_bin_0_drop_rate_mean": 0.12196969696969699, "action_bin_trans_bin_0_drop_rate_std": 0.024218738429008756, "action_bin_trans_bin_0_prob_delta_mean_mean": -0.017464171683051175, "action_bin_trans_bin_0_prob_delta_mean_std": 0.011566807579985917, "action_bin_trans_bin_0_success_delta_mean_mean": -0.1053030303030303, "action_bin_trans_bin_0_success_delta_mean_std": 0.014413861810939732, "action_bin_trans_bin_1_count_mean": 388.6666666666667, "action_bin_trans_bin_1_count_std": 12.39175353029407, "action_bin_trans_bin_1_drop_rate_mean": 0.12033646772852298, "action_bin_trans_bin_1_drop_rate_std": 0.013720020358783488, "action_bin_trans_bin_1_prob_delta_mean_mean": 0.00016739001145992462, "action_bin_trans_bin_1_prob_delta_mean_std": 0.0011143287660981634, "action_bin_trans_bin_1_success_delta_mean_mean": -0.05593398579927078, "action_bin_trans_bin_1_success_delta_mean_std": 0.009354337515458739, "action_bin_trans_bin_2_count_mean": 597.0, "action_bin_trans_bin_2_count_std": 19.949937343260004, "action_bin_trans_bin_2_drop_rate_mean": 0.10472961937267618, "action_bin_trans_bin_2_drop_rate_std": 0.01130578060174296, "action_bin_trans_bin_2_prob_delta_mean_mean": 0.06946160844755211, "action_bin_trans_bin_2_prob_delta_mean_std": 0.005521140772604341, "action_bin_trans_bin_2_success_delta_mean_mean": 0.1157683746879435, "action_bin_trans_bin_2_success_delta_mean_std": 0.025645460314163652, "action_bin_trans_bin_3_count_mean": 292.3333333333333, "action_bin_trans_bin_3_count_std": 11.897712198383164, "action_bin_trans_bin_3_drop_rate_mean": 0.034043075042312264, "action_bin_trans_bin_3_drop_rate_std": 0.004289897738245259, "action_bin_trans_bin_3_prob_delta_mean_mean": 0.17579481909709826, "action_bin_trans_bin_3_prob_delta_mean_std": 0.005624712132983405, "action_bin_trans_bin_3_success_delta_mean_mean": 0.2793901704561506, "action_bin_trans_bin_3_success_delta_mean_std": 0.035813562748876926, "action_dim_0_mean_mean": 0.7473743235721039, "action_dim_0_mean_std": 0.00384631349649342, "action_dim_0_saturation_rate_mean": 0.33666666666666667, "action_dim_0_saturation_rate_std": 0.002377850896280966, "action_dim_0_std_mean": 0.2088044675109674, "action_dim_0_std_std": 0.0020525262200312175, "action_dim_1_mean_mean": -0.10732921867388845, "action_dim_1_mean_std": 0.006143743566069876, "action_dim_1_saturation_rate_mean": 0.07717948717948718, "action_dim_1_saturation_rate_std": 0.006196433832612601, "action_dim_1_std_mean": 0.4844191335352159, "action_dim_1_std_std": 0.010683612945533355, "action_dim_2_mean_mean": 0.08310190335878033, "action_dim_2_mean_std": 0.012684538100538245, "action_dim_2_saturation_rate_mean": 0.1335897435897436, "action_dim_2_saturation_rate_std": 0.006774279391601489, "action_dim_2_std_mean": 0.5629715821499571, "action_dim_2_std_std": 0.005497255079864549, "action_dim_3_mean_mean": 0.39124396397284444, "action_dim_3_mean_std": 0.004778371977881975, "action_dim_3_saturation_rate_mean": 0.02461538461538462, "action_dim_3_saturation_rate_std": 0.0022645540682891922, "action_dim_3_std_mean": 0.221959543255855, "action_dim_3_std_std": 0.0017453618840683137, "action_dim_4_mean_mean": -0.40499699902694636, "action_dim_4_mean_std": 0.0018545519294500606, "action_dim_4_saturation_rate_mean": 0.1346153846153846, "action_dim_4_saturation_rate_std": 0.006185814400930677, "action_dim_4_std_mean": 0.3733014035149285, "action_dim_4_std_std": 0.002336656715763403, "action_dim_5_mean_mean": -0.0324874368216917, "action_dim_5_mean_std": 0.00323360838896328, "action_dim_5_saturation_rate_mean": 0.00717948717948718, "action_dim_5_saturation_rate_std": 0.000725237724293895, "action_dim_5_std_mean": 0.3702717009119703, "action_dim_5_std_std": 0.006576786580936254, "action_rotation_norm_mean_mean": 0.7201990985431852, "action_rotation_norm_mean_std": 0.004008909762347167, "action_rotation_norm_std_mean": 0.3538648042168368, "action_rotation_norm_std_std": 0.006027234691276168, "action_saturation_rate_mean": 0.11897435897435897, "action_saturation_rate_std": 0.003432132137651497, "action_translation_norm_mean_mean": 1.0342196183728154, "action_translation_norm_mean_std": 0.002103999952823672, "action_translation_norm_std_mean": 0.3208069955860246, "action_translation_norm_std_std": 0.004723047331725776, "calibrator_after_brier_mean": 0.19258543401930975, "calibrator_after_brier_std": 0.0010609847242303695, "calibrator_before_brier_vs_legacy_mean": 0.17763223344291493, "calibrator_before_brier_vs_legacy_std": 0.0015463284405348619, "calibrator_neg_prob_delta_degradation_auc_mean": 0.5717138075349452, "calibrator_neg_prob_delta_degradation_auc_std": 0.0315658114597738, "calibrator_prob_after_auc_mean": 0.7097083167382529, "calibrator_prob_after_auc_std": 0.02506237693442019, "calibrator_prob_before_auc_vs_legacy_mean": 0.8620173223811037, "calibrator_prob_before_auc_vs_legacy_std": 0.007016169223295788, "calibrator_prob_delta_recovery_auc_mean": 0.5003729795612851, "calibrator_prob_delta_recovery_auc_std": 0.029288936276980763, "calibrator_raw_logit_after_auc_mean": 0.7097083167382529, "calibrator_raw_logit_after_auc_std": 0.02506237693442019, "calibrator_raw_logit_before_auc_vs_legacy_mean": 0.8620173223811037, "calibrator_raw_logit_before_auc_vs_legacy_std": 0.007016169223295788, "corr_dim_0_negative_recovery_mean": -0.13547621879059257, "corr_dim_0_negative_recovery_std": 0.03566811474176725, "corr_dim_0_positive_drop_mean": 0.060046707674645754, "corr_dim_0_positive_drop_std": 0.0331698822286516, "corr_dim_0_prob_delta_mean": 0.5338095522728064, "corr_dim_0_prob_delta_std": 0.011115792254585335, "corr_dim_1_negative_recovery_mean": -0.2389111616158699, "corr_dim_1_negative_recovery_std": 0.05224112205102062, "corr_dim_1_positive_drop_mean": 0.04389745890578658, "corr_dim_1_positive_drop_std": 0.015236302004610564, "corr_dim_1_prob_delta_mean": 0.5025266922260485, "corr_dim_1_prob_delta_std": 0.013487694618780695, "corr_dim_2_negative_recovery_mean": 0.010713031037756183, "corr_dim_2_negative_recovery_std": 0.013001709903349427, "corr_dim_2_positive_drop_mean": -0.036243207200443966, "corr_dim_2_positive_drop_std": 0.011518163092078275, "corr_dim_2_prob_delta_mean": -0.011348626927194442, "corr_dim_2_prob_delta_std": 0.020775325006005083, "corr_dim_3_negative_recovery_mean": -0.23050131323156378, "corr_dim_3_negative_recovery_std": 0.021409405351452707, "corr_dim_3_positive_drop_mean": 0.007078337002473374, "corr_dim_3_positive_drop_std": 0.05886246524963042, "corr_dim_3_prob_delta_mean": 0.49442137595778446, "corr_dim_3_prob_delta_std": 0.01783253794241535, "corr_dim_4_negative_recovery_mean": 0.1440284237875763, "corr_dim_4_negative_recovery_std": 0.026082598914041927, "corr_dim_4_positive_drop_mean": 0.0018769851490918847, "corr_dim_4_positive_drop_std": 0.020808401553424234, "corr_dim_4_prob_delta_mean": -0.4029861288722074, "corr_dim_4_prob_delta_std": 0.03559829085427217, "corr_dim_5_negative_recovery_mean": -0.2323599406659239, "corr_dim_5_negative_recovery_std": 0.05932381144509948, "corr_dim_5_positive_drop_mean": 0.037364997522200794, "corr_dim_5_positive_drop_std": 0.019820831935791206, "corr_dim_5_prob_delta_mean": 0.47364501141866217, "corr_dim_5_prob_delta_std": 0.013528631668674783, "corr_rotation_norm_prob_delta_mean": 0.5523369228713856, "corr_rotation_norm_prob_delta_std": 0.017089777575753067, "corr_rotation_norm_success_delta_mean": 0.21182980223349793, "corr_rotation_norm_success_delta_std": 0.026755563194032543, "corr_translation_norm_prob_delta_mean": 0.5917931714715836, "corr_translation_norm_prob_delta_std": 0.011836615930928341, "corr_translation_norm_success_delta_mean": 0.2483669814518318, "corr_translation_norm_success_delta_std": 0.03170003777084032, "evaluation_wall_minutes": 47.299231705162676, "experiment_name": "full-sga-gsn-seed8", "label": "full-sga-gsn-seed8", "macro_success_lift_ci95_high": 0.12282051282051283, "macro_success_lift_ci95_low": 0.07230769230769231, "macro_success_lift_mean": 0.09692307692307693, "macro_success_lift_std": 0.014849674087764355, "neg_hold_ci95_high": 0.6152971951734875, "neg_hold_ci95_low": 0.5053679780333146, "neg_hold_mean": 0.5608627716522271, "neg_hold_std": 0.013796371380900271, "pos_drop_ci95_high": 0.18574337941928135, "pos_drop_ci95_low": 0.14065011223571194, "pos_drop_mean": 0.16226287747165766, "pos_drop_std": 0.01320116629274858, "prob_delta_mean_ci95_high": 0.08153336571953848, "prob_delta_mean_ci95_low": 0.0619997069326367, "prob_delta_mean_mean": 0.07111787822452087, "prob_delta_mean_std": 0.0033436907060885113, "source_summary_csv": "/rl-grasp-refine/outputs/unseen_test_formal_paper_seed7_episode/full-sga-gsn-seed8/summary.csv", "t_cover_delta_ci95_high": 0.007238144591790744, "t_cover_delta_ci95_low": -0.020843603606415153, "t_cover_delta_mean": -0.008115947057435615, "t_cover_delta_std": 0.004760167975956908, "t_edge_delta_ci95_high": -0.019990447467397855, "t_edge_delta_ci95_low": -0.052508261395125794, "t_edge_delta_mean": -0.03709511615422506, "t_edge_delta_std": 0.005442015705842136 }, { "across_object_lift_iqr_mean": 0.036666666666666625, "across_object_lift_iqr_std": 0.009428090415820616, "across_object_lift_std_mean": 0.031012713214379917, "across_object_lift_std_std": 0.0012350014372788987, "action_bin_rot_bin_0_count_mean": 1300.0, "action_bin_rot_bin_0_count_std": 0.0, "action_bin_rot_bin_0_drop_rate_mean": 0.038974358974358976, "action_bin_rot_bin_0_drop_rate_std": 0.004411448854893653, "action_bin_rot_bin_0_prob_delta_mean_mean": -0.012355186981927264, "action_bin_rot_bin_0_prob_delta_mean_std": 0.0011275346667699113, "action_bin_rot_bin_0_success_delta_mean_mean": -0.0010256410256410254, "action_bin_rot_bin_0_success_delta_mean_std": 0.00749314830574485, "action_bin_rot_bin_1_count_mean": 0.0, "action_bin_rot_bin_1_count_std": 0.0, "action_bin_rot_bin_1_drop_rate_mean": null, "action_bin_rot_bin_1_drop_rate_std": null, "action_bin_rot_bin_1_prob_delta_mean_mean": null, "action_bin_rot_bin_1_prob_delta_mean_std": null, "action_bin_rot_bin_1_success_delta_mean_mean": null, "action_bin_rot_bin_1_success_delta_mean_std": null, "action_bin_rot_bin_2_count_mean": 0.0, "action_bin_rot_bin_2_count_std": 0.0, "action_bin_rot_bin_2_drop_rate_mean": null, "action_bin_rot_bin_2_drop_rate_std": null, "action_bin_rot_bin_2_prob_delta_mean_mean": null, "action_bin_rot_bin_2_prob_delta_mean_std": null, "action_bin_rot_bin_2_success_delta_mean_mean": null, "action_bin_rot_bin_2_success_delta_mean_std": null, "action_bin_rot_bin_3_count_mean": 0.0, "action_bin_rot_bin_3_count_std": 0.0, "action_bin_rot_bin_3_drop_rate_mean": null, "action_bin_rot_bin_3_drop_rate_std": null, "action_bin_rot_bin_3_prob_delta_mean_mean": null, "action_bin_rot_bin_3_prob_delta_mean_std": null, "action_bin_rot_bin_3_success_delta_mean_mean": null, "action_bin_rot_bin_3_success_delta_mean_std": null, "action_bin_trans_bin_0_count_mean": 1300.0, "action_bin_trans_bin_0_count_std": 0.0, "action_bin_trans_bin_0_drop_rate_mean": 0.038974358974358976, "action_bin_trans_bin_0_drop_rate_std": 0.004411448854893653, "action_bin_trans_bin_0_prob_delta_mean_mean": -0.012355186981927264, "action_bin_trans_bin_0_prob_delta_mean_std": 0.0011275346667699113, "action_bin_trans_bin_0_success_delta_mean_mean": -0.0010256410256410254, "action_bin_trans_bin_0_success_delta_mean_std": 0.00749314830574485, "action_bin_trans_bin_1_count_mean": 0.0, "action_bin_trans_bin_1_count_std": 0.0, "action_bin_trans_bin_1_drop_rate_mean": null, "action_bin_trans_bin_1_drop_rate_std": null, "action_bin_trans_bin_1_prob_delta_mean_mean": null, "action_bin_trans_bin_1_prob_delta_mean_std": null, "action_bin_trans_bin_1_success_delta_mean_mean": null, "action_bin_trans_bin_1_success_delta_mean_std": null, "action_bin_trans_bin_2_count_mean": 0.0, "action_bin_trans_bin_2_count_std": 0.0, "action_bin_trans_bin_2_drop_rate_mean": null, "action_bin_trans_bin_2_drop_rate_std": null, "action_bin_trans_bin_2_prob_delta_mean_mean": null, "action_bin_trans_bin_2_prob_delta_mean_std": null, "action_bin_trans_bin_2_success_delta_mean_mean": null, "action_bin_trans_bin_2_success_delta_mean_std": null, "action_bin_trans_bin_3_count_mean": 0.0, "action_bin_trans_bin_3_count_std": 0.0, "action_bin_trans_bin_3_drop_rate_mean": null, "action_bin_trans_bin_3_drop_rate_std": null, "action_bin_trans_bin_3_prob_delta_mean_mean": null, "action_bin_trans_bin_3_prob_delta_mean_std": null, "action_bin_trans_bin_3_success_delta_mean_mean": null, "action_bin_trans_bin_3_success_delta_mean_std": null, "action_dim_0_mean_mean": 0.0, "action_dim_0_mean_std": 0.0, "action_dim_0_saturation_rate_mean": 0.0, "action_dim_0_saturation_rate_std": 0.0, "action_dim_0_std_mean": 0.0, "action_dim_0_std_std": 0.0, "action_dim_1_mean_mean": 0.0, "action_dim_1_mean_std": 0.0, "action_dim_1_saturation_rate_mean": 0.0, "action_dim_1_saturation_rate_std": 0.0, "action_dim_1_std_mean": 0.0, "action_dim_1_std_std": 0.0, "action_dim_2_mean_mean": 0.0, "action_dim_2_mean_std": 0.0, "action_dim_2_saturation_rate_mean": 0.0, "action_dim_2_saturation_rate_std": 0.0, "action_dim_2_std_mean": 0.0, "action_dim_2_std_std": 0.0, "action_dim_3_mean_mean": 0.0, "action_dim_3_mean_std": 0.0, "action_dim_3_saturation_rate_mean": 0.0, "action_dim_3_saturation_rate_std": 0.0, "action_dim_3_std_mean": 0.0, "action_dim_3_std_std": 0.0, "action_dim_4_mean_mean": 0.0, "action_dim_4_mean_std": 0.0, "action_dim_4_saturation_rate_mean": 0.0, "action_dim_4_saturation_rate_std": 0.0, "action_dim_4_std_mean": 0.0, "action_dim_4_std_std": 0.0, "action_dim_5_mean_mean": 0.0, "action_dim_5_mean_std": 0.0, "action_dim_5_saturation_rate_mean": 0.0, "action_dim_5_saturation_rate_std": 0.0, "action_dim_5_std_mean": 0.0, "action_dim_5_std_std": 0.0, "action_rotation_norm_mean_mean": 0.0, "action_rotation_norm_mean_std": 0.0, "action_rotation_norm_std_mean": 0.0, "action_rotation_norm_std_std": 0.0, "action_saturation_rate_mean": 0.0, "action_saturation_rate_std": 0.0, "action_translation_norm_mean_mean": 0.0, "action_translation_norm_mean_std": 0.0, "action_translation_norm_std_mean": 0.0, "action_translation_norm_std_std": 0.0, "calibrator_after_brier_mean": 0.19438051104089457, "calibrator_after_brier_std": 0.001424579650315923, "calibrator_before_brier_vs_legacy_mean": 0.19338580381636997, "calibrator_before_brier_vs_legacy_std": 0.0008461882395214853, "calibrator_neg_prob_delta_degradation_auc_mean": 0.5215535552049482, "calibrator_neg_prob_delta_degradation_auc_std": 0.017389486048094642, "calibrator_prob_after_auc_mean": 0.8598128270185416, "calibrator_prob_after_auc_std": 0.007038308252212574, "calibrator_prob_before_auc_vs_legacy_mean": 0.8618807850337692, "calibrator_prob_before_auc_vs_legacy_std": 0.005728139636066861, "calibrator_prob_delta_recovery_auc_mean": 0.518630817084915, "calibrator_prob_delta_recovery_auc_std": 0.04276878774618757, "calibrator_raw_logit_after_auc_mean": 0.8598128270185416, "calibrator_raw_logit_after_auc_std": 0.007038308252212574, "calibrator_raw_logit_before_auc_vs_legacy_mean": 0.8618807850337692, "calibrator_raw_logit_before_auc_vs_legacy_std": 0.005728139636066861, "corr_dim_0_negative_recovery_mean": null, "corr_dim_0_negative_recovery_std": null, "corr_dim_0_positive_drop_mean": null, "corr_dim_0_positive_drop_std": null, "corr_dim_0_prob_delta_mean": null, "corr_dim_0_prob_delta_std": null, "corr_dim_1_negative_recovery_mean": null, "corr_dim_1_negative_recovery_std": null, "corr_dim_1_positive_drop_mean": null, "corr_dim_1_positive_drop_std": null, "corr_dim_1_prob_delta_mean": null, "corr_dim_1_prob_delta_std": null, "corr_dim_2_negative_recovery_mean": null, "corr_dim_2_negative_recovery_std": null, "corr_dim_2_positive_drop_mean": null, "corr_dim_2_positive_drop_std": null, "corr_dim_2_prob_delta_mean": null, "corr_dim_2_prob_delta_std": null, "corr_dim_3_negative_recovery_mean": null, "corr_dim_3_negative_recovery_std": null, "corr_dim_3_positive_drop_mean": null, "corr_dim_3_positive_drop_std": null, "corr_dim_3_prob_delta_mean": null, "corr_dim_3_prob_delta_std": null, "corr_dim_4_negative_recovery_mean": null, "corr_dim_4_negative_recovery_std": null, "corr_dim_4_positive_drop_mean": null, "corr_dim_4_positive_drop_std": null, "corr_dim_4_prob_delta_mean": null, "corr_dim_4_prob_delta_std": null, "corr_dim_5_negative_recovery_mean": null, "corr_dim_5_negative_recovery_std": null, "corr_dim_5_positive_drop_mean": null, "corr_dim_5_positive_drop_std": null, "corr_dim_5_prob_delta_mean": null, "corr_dim_5_prob_delta_std": null, "corr_rotation_norm_prob_delta_mean": null, "corr_rotation_norm_prob_delta_std": null, "corr_rotation_norm_success_delta_mean": null, "corr_rotation_norm_success_delta_std": null, "corr_translation_norm_prob_delta_mean": null, "corr_translation_norm_prob_delta_std": null, "corr_translation_norm_success_delta_mean": null, "corr_translation_norm_success_delta_std": null, "evaluation_wall_minutes": 44.83580386828786, "experiment_name": "no-action", "label": "no-action", "macro_success_lift_ci95_high": 0.008461538461538463, "macro_success_lift_ci95_low": -0.010769230769230781, "macro_success_lift_mean": -0.0010256410256410246, "macro_success_lift_std": 0.007493148305744844, "neg_hold_ci95_high": 0.14256207141074434, "neg_hold_ci95_low": 0.09443550698142744, "neg_hold_mean": 0.11824745529719692, "neg_hold_std": 0.01386845164857522, "pos_drop_ci95_high": 0.07973806059012169, "pos_drop_ci95_low": 0.055443235048624716, "pos_drop_mean": 0.06768525244041983, "pos_drop_std": 0.011217829141652338, "prob_delta_mean_ci95_high": -0.010424051277465438, "prob_delta_mean_ci95_low": -0.014188788084868845, "prob_delta_mean_mean": -0.012355186981927266, "prob_delta_mean_std": 0.0011275346667699113, "source_summary_csv": "/rl-grasp-refine/outputs/unseen_test_formal_paper_seed7_episode/no-action/summary.csv", "t_cover_delta_ci95_high": -0.04426919671981825, "t_cover_delta_ci95_low": -0.05133538704995945, "t_cover_delta_mean": -0.047632351768154946, "t_cover_delta_std": 0.0019143588699239216, "t_edge_delta_ci95_high": -0.0909381242775812, "t_edge_delta_ci95_low": -0.10465632955580488, "t_edge_delta_mean": -0.09784604655701118, "t_edge_delta_std": 0.0013233131199331418 }, { "across_object_lift_iqr_mean": 0.10666666666666665, "across_object_lift_iqr_std": 0.024944382578492966, "across_object_lift_std_mean": 0.08209453775578039, "across_object_lift_std_std": 0.01600461246552532, "action_bin_rot_bin_0_count_mean": 57.666666666666664, "action_bin_rot_bin_0_count_std": 3.8586123009300755, "action_bin_rot_bin_0_drop_rate_mean": 0.1902557319223986, "action_bin_rot_bin_0_drop_rate_std": 0.034291050691173226, "action_bin_rot_bin_0_prob_delta_mean_mean": -0.0006586056754724709, "action_bin_rot_bin_0_prob_delta_mean_std": 0.009194986017672022, "action_bin_rot_bin_0_success_delta_mean_mean": -0.11067019400352733, "action_bin_rot_bin_0_success_delta_mean_std": 0.04806198723961395, "action_bin_rot_bin_1_count_mean": 383.6666666666667, "action_bin_rot_bin_1_count_std": 2.0548046676563256, "action_bin_rot_bin_1_drop_rate_mean": 0.216312673108219, "action_bin_rot_bin_1_drop_rate_std": 0.004601906788434182, "action_bin_rot_bin_1_prob_delta_mean_mean": -0.01040761341170286, "action_bin_rot_bin_1_prob_delta_mean_std": 0.0032627166302414555, "action_bin_rot_bin_1_success_delta_mean_mean": -0.1372358258083672, "action_bin_rot_bin_1_success_delta_mean_std": 0.006746233057080026, "action_bin_rot_bin_2_count_mean": 713.0, "action_bin_rot_bin_2_count_std": 5.0990195135927845, "action_bin_rot_bin_2_drop_rate_mean": 0.1861522068904868, "action_bin_rot_bin_2_drop_rate_std": 0.011768225655342869, "action_bin_rot_bin_2_prob_delta_mean_mean": -0.008668235447252289, "action_bin_rot_bin_2_prob_delta_mean_std": 0.001183581969158079, "action_bin_rot_bin_2_success_delta_mean_mean": -0.09598480355777539, "action_bin_rot_bin_2_success_delta_mean_std": 0.021303572516972683, "action_bin_rot_bin_3_count_mean": 145.66666666666666, "action_bin_rot_bin_3_count_std": 3.299831645537222, "action_bin_rot_bin_3_drop_rate_mean": 0.21530624359181913, "action_bin_rot_bin_3_drop_rate_std": 0.03334379459023273, "action_bin_rot_bin_3_prob_delta_mean_mean": -0.009519226088562788, "action_bin_rot_bin_3_prob_delta_mean_std": 0.004307804129806013, "action_bin_rot_bin_3_success_delta_mean_mean": -0.1378371377691436, "action_bin_rot_bin_3_success_delta_mean_std": 0.05610328318499789, "action_bin_trans_bin_0_count_mean": 60.333333333333336, "action_bin_trans_bin_0_count_std": 6.128258770283412, "action_bin_trans_bin_0_drop_rate_mean": 0.09770008632383771, "action_bin_trans_bin_0_drop_rate_std": 0.03895570269853402, "action_bin_trans_bin_0_prob_delta_mean_mean": -0.012778746420450543, "action_bin_trans_bin_0_prob_delta_mean_std": 0.005091762331156924, "action_bin_trans_bin_0_success_delta_mean_mean": -0.03406091996547047, "action_bin_trans_bin_0_success_delta_mean_std": 0.06664377352851762, "action_bin_trans_bin_1_count_mean": 389.0, "action_bin_trans_bin_1_count_std": 5.715476066494082, "action_bin_trans_bin_1_drop_rate_mean": 0.14991187232539963, "action_bin_trans_bin_1_drop_rate_std": 0.015613058908673406, "action_bin_trans_bin_1_prob_delta_mean_mean": -0.013295022153424213, "action_bin_trans_bin_1_prob_delta_mean_std": 0.0015086227899343157, "action_bin_trans_bin_1_success_delta_mean_mean": -0.07280297802386639, "action_bin_trans_bin_1_success_delta_mean_std": 0.014692274065081149, "action_bin_trans_bin_2_count_mean": 707.0, "action_bin_trans_bin_2_count_std": 1.632993161855452, "action_bin_trans_bin_2_drop_rate_mean": 0.22253640165462138, "action_bin_trans_bin_2_drop_rate_std": 0.002334349918685093, "action_bin_trans_bin_2_prob_delta_mean_mean": -0.007332928847468474, "action_bin_trans_bin_2_prob_delta_mean_std": 0.0006824546049679038, "action_bin_trans_bin_2_success_delta_mean_mean": -0.1343686066490136, "action_bin_trans_bin_2_success_delta_mean_std": 0.006062795278583714, "action_bin_trans_bin_3_count_mean": 143.66666666666666, "action_bin_trans_bin_3_count_std": 4.109609335312651, "action_bin_trans_bin_3_drop_rate_mean": 0.25537885737779514, "action_bin_trans_bin_3_drop_rate_std": 0.012953083807695867, "action_bin_trans_bin_3_prob_delta_mean_mean": -0.0027246129759339204, "action_bin_trans_bin_3_prob_delta_mean_std": 0.004901730185057656, "action_bin_trans_bin_3_success_delta_mean_mean": -0.15525245359632886, "action_bin_trans_bin_3_success_delta_mean_std": 0.03048468332735315, "action_dim_0_mean_mean": 0.006689010741463743, "action_dim_0_mean_std": 0.012774980799600908, "action_dim_0_saturation_rate_mean": 0.10435897435897436, "action_dim_0_saturation_rate_std": 0.006715795309594255, "action_dim_0_std_mean": 0.5786235544478223, "action_dim_0_std_std": 0.0018898364469294767, "action_dim_1_mean_mean": -0.006383694069827937, "action_dim_1_mean_std": 0.011835164823950122, "action_dim_1_saturation_rate_mean": 0.10230769230769231, "action_dim_1_saturation_rate_std": 0.00062807429302133, "action_dim_1_std_mean": 0.5786814362546956, "action_dim_1_std_std": 0.0022205967324948174, "action_dim_2_mean_mean": 0.00591551667166641, "action_dim_2_mean_std": 0.00724018003488999, "action_dim_2_saturation_rate_mean": 0.09487179487179487, "action_dim_2_saturation_rate_std": 0.006946778043150806, "action_dim_2_std_mean": 0.5714148830073446, "action_dim_2_std_std": 0.009118306146917401, "action_dim_3_mean_mean": -0.0014333630625706292, "action_dim_3_mean_std": 0.00568304667368754, "action_dim_3_saturation_rate_mean": 0.10384615384615385, "action_dim_3_saturation_rate_std": 0.0057563959796522145, "action_dim_3_std_mean": 0.5796320597905038, "action_dim_3_std_std": 0.0028881756180852195, "action_dim_4_mean_mean": 0.002886493541322685, "action_dim_4_mean_std": 0.0052796740203824135, "action_dim_4_saturation_rate_mean": 0.10435897435897436, "action_dim_4_saturation_rate_std": 0.002018976377951747, "action_dim_4_std_mean": 0.5772148667097872, "action_dim_4_std_std": 0.004778069790187221, "action_dim_5_mean_mean": 0.00047351919263840153, "action_dim_5_mean_std": 0.026279655606128955, "action_dim_5_saturation_rate_mean": 0.10564102564102563, "action_dim_5_saturation_rate_std": 0.00817301407718422, "action_dim_5_std_mean": 0.5789319115949626, "action_dim_5_std_std": 0.003690536628943503, "action_rotation_norm_mean_mean": 0.9619135429555641, "action_rotation_norm_mean_std": 0.0033069261213848223, "action_rotation_norm_std_mean": 0.282536496852079, "action_rotation_norm_std_std": 0.0017910490382142916, "action_saturation_rate_mean": 0.10256410256410257, "action_saturation_rate_std": 0.002902838922690892, "action_translation_norm_mean_mean": 0.9571313643480748, "action_translation_norm_mean_std": 0.004942988586318917, "action_translation_norm_std_mean": 0.28395225839728483, "action_translation_norm_std_std": 0.0025395678012702866, "calibrator_after_brier_mean": 0.2167926450035013, "calibrator_after_brier_std": 0.0018353011556495114, "calibrator_before_brier_vs_legacy_mean": 0.1933407271007843, "calibrator_before_brier_vs_legacy_std": 0.0005150504147324581, "calibrator_neg_prob_delta_degradation_auc_mean": 0.5965328149626652, "calibrator_neg_prob_delta_degradation_auc_std": 0.0017363610311709765, "calibrator_prob_after_auc_mean": 0.7755454123541315, "calibrator_prob_after_auc_std": 0.007098696866181568, "calibrator_prob_before_auc_vs_legacy_mean": 0.8618829528071471, "calibrator_prob_before_auc_vs_legacy_std": 0.003695398658339251, "calibrator_prob_delta_recovery_auc_mean": 0.5914983910238725, "calibrator_prob_delta_recovery_auc_std": 0.014090496824556056, "calibrator_raw_logit_after_auc_mean": 0.7755454123541315, "calibrator_raw_logit_after_auc_std": 0.007098696866181568, "calibrator_raw_logit_before_auc_vs_legacy_mean": 0.8618829528071471, "calibrator_raw_logit_before_auc_vs_legacy_std": 0.003695398658339251, "corr_dim_0_negative_recovery_mean": 0.29968951233616753, "corr_dim_0_negative_recovery_std": 0.028959830347118243, "corr_dim_0_positive_drop_mean": -0.16126144954728364, "corr_dim_0_positive_drop_std": 0.03051776110445836, "corr_dim_0_prob_delta_mean": 0.344259987012707, "corr_dim_0_prob_delta_std": 0.01996361342720525, "corr_dim_1_negative_recovery_mean": -0.0031501235220138114, "corr_dim_1_negative_recovery_std": 0.03227017195914447, "corr_dim_1_positive_drop_mean": 0.006705651593270194, "corr_dim_1_positive_drop_std": 0.025783235096578695, "corr_dim_1_prob_delta_mean": 0.018674080014359162, "corr_dim_1_prob_delta_std": 0.010819823202306565, "corr_dim_2_negative_recovery_mean": -0.0009835124548082414, "corr_dim_2_negative_recovery_std": 0.023166915357807908, "corr_dim_2_positive_drop_mean": 0.011299027226331517, "corr_dim_2_positive_drop_std": 0.02526633264436376, "corr_dim_2_prob_delta_mean": -0.017002238324373784, "corr_dim_2_prob_delta_std": 0.039765484993303644, "corr_dim_3_negative_recovery_mean": 0.03155753547404259, "corr_dim_3_negative_recovery_std": 0.01641398454873658, "corr_dim_3_positive_drop_mean": -0.02969710456305609, "corr_dim_3_positive_drop_std": 0.014739092080296654, "corr_dim_3_prob_delta_mean": -0.041237316935724126, "corr_dim_3_prob_delta_std": 0.0373455760373439, "corr_dim_4_negative_recovery_mean": 0.01926673771150687, "corr_dim_4_negative_recovery_std": 0.0449074601180045, "corr_dim_4_positive_drop_mean": 0.005188361421214028, "corr_dim_4_positive_drop_std": 0.027338068390037537, "corr_dim_4_prob_delta_mean": -0.010507051644916277, "corr_dim_4_prob_delta_std": 0.030058832663348946, "corr_dim_5_negative_recovery_mean": 0.020224847799740462, "corr_dim_5_negative_recovery_std": 0.04399853367123815, "corr_dim_5_positive_drop_mean": -0.021493827511801927, "corr_dim_5_positive_drop_std": 0.019772857776786393, "corr_dim_5_prob_delta_mean": 0.04011502336403408, "corr_dim_5_prob_delta_std": 0.010724378754432442, "corr_rotation_norm_prob_delta_mean": 0.001367860845185234, "corr_rotation_norm_prob_delta_std": 0.019271495868651095, "corr_rotation_norm_success_delta_mean": 0.0183162191329268, "corr_rotation_norm_success_delta_std": 0.015550144638304607, "corr_translation_norm_prob_delta_mean": 0.04265478546489976, "corr_translation_norm_prob_delta_std": 0.005918139812426922, "corr_translation_norm_success_delta_mean": -0.06886672570231019, "corr_translation_norm_success_delta_std": 0.014015916981303972, "evaluation_wall_minutes": 44.14993973591675, "experiment_name": "rand-action", "label": "rand-action", "macro_success_lift_ci95_high": -0.08692307692307694, "macro_success_lift_ci95_low": -0.13846153846153844, "macro_success_lift_mean": -0.11333333333333334, "macro_success_lift_std": 0.011586793361738457, "neg_hold_ci95_high": 0.28735798551273134, "neg_hold_ci95_low": 0.22022594491030137, "neg_hold_mean": 0.25359166342128225, "neg_hold_std": 0.01859678256799903, "pos_drop_ci95_high": 0.3771335875393348, "pos_drop_ci95_low": 0.32198323765719766, "pos_drop_mean": 0.3487118440140831, "pos_drop_std": 0.008202846299930405, "prob_delta_mean_ci95_high": -0.0050515780113822295, "prob_delta_mean_ci95_low": -0.012338978535163801, "prob_delta_mean_mean": -0.008880441190868659, "prob_delta_mean_std": 0.0006405991612058615, "source_summary_csv": "/rl-grasp-refine/outputs/unseen_test_formal_paper_seed7_episode/rand-action/summary.csv", "t_cover_delta_ci95_high": -0.029086887438115095, "t_cover_delta_ci95_low": -0.04274920964474885, "t_cover_delta_mean": -0.035898255547274924, "t_cover_delta_std": 0.004390047746185666, "t_edge_delta_ci95_high": -0.09130570617525331, "t_edge_delta_ci95_low": -0.11255003372392579, "t_edge_delta_mean": -0.10212315514105379, "t_edge_delta_std": 0.004848150971582201 }, { "across_object_lift_iqr_mean": 0.11000000000000003, "across_object_lift_iqr_std": 0.048989794855663606, "across_object_lift_std_mean": 0.07867266536775706, "across_object_lift_std_std": 0.010840437121726991, "action_bin_rot_bin_0_count_mean": 547.6666666666666, "action_bin_rot_bin_0_count_std": 13.572848714334889, "action_bin_rot_bin_0_drop_rate_mean": 0.1010942360364258, "action_bin_rot_bin_0_drop_rate_std": 0.002535443573253168, "action_bin_rot_bin_0_prob_delta_mean_mean": -0.00244055236453145, "action_bin_rot_bin_0_prob_delta_mean_std": 0.008278296465443875, "action_bin_rot_bin_0_success_delta_mean_mean": -0.020172724936959662, "action_bin_rot_bin_0_success_delta_mean_std": 0.01124053133496951, "action_bin_rot_bin_1_count_mean": 392.3333333333333, "action_bin_rot_bin_1_count_std": 15.4344492037203, "action_bin_rot_bin_1_drop_rate_mean": 0.09828350642650542, "action_bin_rot_bin_1_drop_rate_std": 0.0069591202890725845, "action_bin_rot_bin_1_prob_delta_mean_mean": 0.19083685177428547, "action_bin_rot_bin_1_prob_delta_mean_std": 0.0038240458637030395, "action_bin_rot_bin_1_success_delta_mean_mean": 0.1315729637477403, "action_bin_rot_bin_1_success_delta_mean_std": 0.01887271764938763, "action_bin_rot_bin_2_count_mean": 332.6666666666667, "action_bin_rot_bin_2_count_std": 10.274023338281626, "action_bin_rot_bin_2_drop_rate_mean": 0.05487324530496542, "action_bin_rot_bin_2_drop_rate_std": 0.011133049193178703, "action_bin_rot_bin_2_prob_delta_mean_mean": 0.33305453126718404, "action_bin_rot_bin_2_prob_delta_mean_std": 0.009922632014572325, "action_bin_rot_bin_2_success_delta_mean_mean": 0.2443795020657541, "action_bin_rot_bin_2_success_delta_mean_std": 0.03426262957590373, "action_bin_rot_bin_3_count_mean": 27.333333333333332, "action_bin_rot_bin_3_count_std": 5.2493385826745405, "action_bin_rot_bin_3_drop_rate_mean": 0.010416666666666666, "action_bin_rot_bin_3_drop_rate_std": 0.01473139127471974, "action_bin_rot_bin_3_prob_delta_mean_mean": 0.18903940252391047, "action_bin_rot_bin_3_prob_delta_mean_std": 0.013559627182995474, "action_bin_rot_bin_3_success_delta_mean_mean": 0.31527777777777777, "action_bin_rot_bin_3_success_delta_mean_std": 0.013749298523071754, "action_bin_trans_bin_0_count_mean": 169.66666666666666, "action_bin_trans_bin_0_count_std": 15.691469727919761, "action_bin_trans_bin_0_drop_rate_mean": 0.06798029195560967, "action_bin_trans_bin_0_drop_rate_std": 0.021867257733131702, "action_bin_trans_bin_0_prob_delta_mean_mean": -0.045746334623069845, "action_bin_trans_bin_0_prob_delta_mean_std": 0.007136992585080743, "action_bin_trans_bin_0_success_delta_mean_mean": -0.05385025562383855, "action_bin_trans_bin_0_success_delta_mean_std": 0.018674159509356213, "action_bin_trans_bin_1_count_mean": 397.3333333333333, "action_bin_trans_bin_1_count_std": 20.741798914805393, "action_bin_trans_bin_1_drop_rate_mean": 0.12186048965808359, "action_bin_trans_bin_1_drop_rate_std": 0.00573250209471417, "action_bin_trans_bin_1_prob_delta_mean_mean": 0.0048403505986495485, "action_bin_trans_bin_1_prob_delta_mean_std": 0.005435347653381893, "action_bin_trans_bin_1_success_delta_mean_mean": -0.019769241592795833, "action_bin_trans_bin_1_success_delta_mean_std": 0.009553147487585912, "action_bin_trans_bin_2_count_mean": 541.6666666666666, "action_bin_trans_bin_2_count_std": 21.044925490219462, "action_bin_trans_bin_2_drop_rate_mean": 0.08173144876770784, "action_bin_trans_bin_2_drop_rate_std": 0.002952876019485394, "action_bin_trans_bin_2_prob_delta_mean_mean": 0.23694916809060704, "action_bin_trans_bin_2_prob_delta_mean_std": 0.004827315628170542, "action_bin_trans_bin_2_success_delta_mean_mean": 0.1731996811883868, "action_bin_trans_bin_2_success_delta_mean_std": 0.009880101097924518, "action_bin_trans_bin_3_count_mean": 191.33333333333334, "action_bin_trans_bin_3_count_std": 8.65383665716478, "action_bin_trans_bin_3_drop_rate_mean": 0.044566316374374526, "action_bin_trans_bin_3_drop_rate_std": 0.018073663106240374, "action_bin_trans_bin_3_prob_delta_mean_mean": 0.35088735966796564, "action_bin_trans_bin_3_prob_delta_mean_std": 0.0033360236098666446, "action_bin_trans_bin_3_success_delta_mean_mean": 0.2788774893523173, "action_bin_trans_bin_3_success_delta_mean_std": 0.042634356076302594, "action_dim_0_mean_mean": 0.7259989651871056, "action_dim_0_mean_std": 0.004824782082083792, "action_dim_0_saturation_rate_mean": 0.4069230769230769, "action_dim_0_saturation_rate_std": 0.006793662204867587, "action_dim_0_std_mean": 0.2631012808349966, "action_dim_0_std_std": 0.0037052017937269244, "action_dim_1_mean_mean": 0.1423534373746728, "action_dim_1_mean_std": 0.003670000824529201, "action_dim_1_saturation_rate_mean": 0.0, "action_dim_1_saturation_rate_std": 0.0, "action_dim_1_std_mean": 0.19378864755910033, "action_dim_1_std_std": 0.0017025255566052542, "action_dim_2_mean_mean": 0.027373457916382442, "action_dim_2_mean_std": 0.011710526513982373, "action_dim_2_saturation_rate_mean": 0.09256410256410257, "action_dim_2_saturation_rate_std": 0.00785302196467362, "action_dim_2_std_mean": 0.5254085214696325, "action_dim_2_std_std": 0.0028875581847987437, "action_dim_3_mean_mean": -0.04337513621392039, "action_dim_3_mean_std": 0.004085565103489755, "action_dim_3_saturation_rate_mean": 0.0010256410256410256, "action_dim_3_saturation_rate_std": 0.0007252377242938949, "action_dim_3_std_mean": 0.2778867007593743, "action_dim_3_std_std": 0.002181356010695416, "action_dim_4_mean_mean": 0.331556875008242, "action_dim_4_mean_std": 0.004567617011618664, "action_dim_4_saturation_rate_mean": 0.01769230769230769, "action_dim_4_saturation_rate_std": 0.0016617283842071439, "action_dim_4_std_mean": 0.2610038294282741, "action_dim_4_std_std": 0.0035562629942868127, "action_dim_5_mean_mean": 0.21310133691898023, "action_dim_5_mean_std": 0.009118145855236534, "action_dim_5_saturation_rate_mean": 0.007948717948717949, "action_dim_5_saturation_rate_std": 0.0007252377242938947, "action_dim_5_std_mean": 0.4104498337880929, "action_dim_5_std_std": 0.001736919663357955, "action_rotation_norm_mean_mean": 0.5907661476046868, "action_rotation_norm_mean_std": 0.006431317716447425, "action_rotation_norm_std_mean": 0.34945623047596136, "action_rotation_norm_std_std": 0.0034763249609018085, "action_saturation_rate_mean": 0.0876923076923077, "action_saturation_rate_std": 0.002670854699487139, "action_translation_norm_mean_mean": 0.8997535367139348, "action_translation_norm_mean_std": 0.006081959348304921, "action_translation_norm_std_mean": 0.34859525614725834, "action_translation_norm_std_std": 0.002349981769760161, "calibrator_after_brier_mean": 0.17636842729139993, "calibrator_after_brier_std": 0.004399571439106009, "calibrator_before_brier_vs_legacy_mean": 0.14790553905878676, "calibrator_before_brier_vs_legacy_std": 0.004896152827920994, "calibrator_neg_prob_delta_degradation_auc_mean": 0.5465065662406575, "calibrator_neg_prob_delta_degradation_auc_std": 0.030492877823745997, "calibrator_prob_after_auc_mean": 0.7555511567110464, "calibrator_prob_after_auc_std": 0.013706648507649553, "calibrator_prob_before_auc_vs_legacy_mean": 0.8624161856036535, "calibrator_prob_before_auc_vs_legacy_std": 0.00674421230760331, "calibrator_prob_delta_recovery_auc_mean": 0.5830571474171197, "calibrator_prob_delta_recovery_auc_std": 0.012962032157051528, "calibrator_raw_logit_after_auc_mean": 0.7555511567110464, "calibrator_raw_logit_after_auc_std": 0.013706648507649553, "calibrator_raw_logit_before_auc_vs_legacy_mean": 0.8624161856036535, "calibrator_raw_logit_before_auc_vs_legacy_std": 0.00674421230760331, "corr_dim_0_negative_recovery_mean": -0.11577357137614801, "corr_dim_0_negative_recovery_std": 0.017165530589252458, "corr_dim_0_positive_drop_mean": 0.11199372939885271, "corr_dim_0_positive_drop_std": 0.025015224143875834, "corr_dim_0_prob_delta_mean": 0.5211428476796559, "corr_dim_0_prob_delta_std": 0.009582100126699028, "corr_dim_1_negative_recovery_mean": 0.0488470631432737, "corr_dim_1_negative_recovery_std": 0.03512197465935067, "corr_dim_1_positive_drop_mean": 0.05258153592638596, "corr_dim_1_positive_drop_std": 0.06453600785783101, "corr_dim_1_prob_delta_mean": 0.3845541246686401, "corr_dim_1_prob_delta_std": 0.018777495564928987, "corr_dim_2_negative_recovery_mean": -0.09082790467613093, "corr_dim_2_negative_recovery_std": 0.06952105634789689, "corr_dim_2_positive_drop_mean": 0.0067364395791760815, "corr_dim_2_positive_drop_std": 0.04819223865378756, "corr_dim_2_prob_delta_mean": 0.19317127425563296, "corr_dim_2_prob_delta_std": 0.039745066612263354, "corr_dim_3_negative_recovery_mean": 0.08253658300545783, "corr_dim_3_negative_recovery_std": 0.05977607162585835, "corr_dim_3_positive_drop_mean": 0.011667174079827224, "corr_dim_3_positive_drop_std": 0.03824204638436238, "corr_dim_3_prob_delta_mean": 0.04886411306174138, "corr_dim_3_prob_delta_std": 0.026483480958236782, "corr_dim_4_negative_recovery_mean": -0.23062860774829433, "corr_dim_4_negative_recovery_std": 0.06475364750559931, "corr_dim_4_positive_drop_mean": 0.1461130976386696, "corr_dim_4_positive_drop_std": 0.055375671928577674, "corr_dim_4_prob_delta_mean": 0.5086915820713843, "corr_dim_4_prob_delta_std": 0.007649536522766335, "corr_dim_5_negative_recovery_mean": -0.08080344843388543, "corr_dim_5_negative_recovery_std": 0.05135314097774648, "corr_dim_5_positive_drop_mean": 0.034809914028241074, "corr_dim_5_positive_drop_std": 0.04989273466473564, "corr_dim_5_prob_delta_mean": 0.3866995047407773, "corr_dim_5_prob_delta_std": 0.03388322477260128, "corr_rotation_norm_prob_delta_mean": 0.519304605401714, "corr_rotation_norm_prob_delta_std": 0.013466015072129378, "corr_rotation_norm_success_delta_mean": 0.2279222219813121, "corr_rotation_norm_success_delta_std": 0.03606376527287688, "corr_translation_norm_prob_delta_mean": 0.5560888259791087, "corr_translation_norm_prob_delta_std": 0.006854981947782115, "corr_translation_norm_success_delta_mean": 0.23462445675288957, "corr_translation_norm_success_delta_std": 0.020473786287991864, "evaluation_wall_minutes": 47.39192778449505, "experiment_name": "vanilla", "label": "vanilla", "macro_success_lift_ci95_high": 0.1253846153846154, "macro_success_lift_ci95_low": 0.07615384615384614, "macro_success_lift_mean": 0.09999999999999999, "macro_success_lift_std": 0.0033234567684143038, "neg_hold_ci95_high": 0.6058522319242963, "neg_hold_ci95_low": 0.49279638219104405, "neg_hold_mean": 0.5483742849285207, "neg_hold_std": 0.016743546878540377, "pos_drop_ci95_high": 0.17688951545379025, "pos_drop_ci95_low": 0.13112079890055175, "pos_drop_mean": 0.15261387402149754, "pos_drop_std": 0.003397848959350471, "prob_delta_mean_ci95_high": 0.16017290287949273, "prob_delta_mean_ci95_low": 0.132209413266326, "prob_delta_mean_mean": 0.14576396324894023, "prob_delta_mean_std": 0.001996138508535177, "source_summary_csv": "/rl-grasp-refine/outputs/unseen_test_formal_paper_seed7_episode/vanilla/summary.csv", "t_cover_delta_ci95_high": -0.02123614097433556, "t_cover_delta_ci95_low": -0.03730829030969055, "t_cover_delta_mean": -0.029887289578148114, "t_cover_delta_std": 0.003080679170822434, "t_edge_delta_ci95_high": -0.04734404831100032, "t_edge_delta_ci95_low": -0.07619690787039758, "t_edge_delta_mean": -0.062301181548525784, "t_edge_delta_std": 0.006081149307910554 }, { "across_object_lift_iqr_mean": 0.15, "across_object_lift_iqr_std": 0.04082482904638634, "across_object_lift_std_mean": 0.09371197862587777, "across_object_lift_std_std": 0.01844865931347976, "action_bin_rot_bin_0_count_mean": 115.0, "action_bin_rot_bin_0_count_std": 17.45470328211473, "action_bin_rot_bin_0_drop_rate_mean": 0.08868817885211328, "action_bin_rot_bin_0_drop_rate_std": 0.02010155338374227, "action_bin_rot_bin_0_prob_delta_mean_mean": 0.14395979677805024, "action_bin_rot_bin_0_prob_delta_mean_std": 0.03896194495566781, "action_bin_rot_bin_0_success_delta_mean_mean": 0.0882742030283014, "action_bin_rot_bin_0_success_delta_mean_std": 0.009542326042111337, "action_bin_rot_bin_1_count_mean": 365.0, "action_bin_rot_bin_1_count_std": 28.577380332470412, "action_bin_rot_bin_1_drop_rate_mean": 0.10586066140860662, "action_bin_rot_bin_1_drop_rate_std": 0.011162266116325868, "action_bin_rot_bin_1_prob_delta_mean_mean": 0.13648834679536223, "action_bin_rot_bin_1_prob_delta_mean_std": 0.009261313745963125, "action_bin_rot_bin_1_success_delta_mean_mean": 0.07094506710945069, "action_bin_rot_bin_1_success_delta_mean_std": 0.020610274086307098, "action_bin_rot_bin_2_count_mean": 769.3333333333334, "action_bin_rot_bin_2_count_std": 12.472191289246473, "action_bin_rot_bin_2_drop_rate_mean": 0.11179171917018967, "action_bin_rot_bin_2_drop_rate_std": 0.0012742177630668458, "action_bin_rot_bin_2_prob_delta_mean_mean": 0.19047659022813104, "action_bin_rot_bin_2_prob_delta_mean_std": 0.009392805697642222, "action_bin_rot_bin_2_success_delta_mean_mean": 0.11048355582102153, "action_bin_rot_bin_2_success_delta_mean_std": 0.019373673978823513, "action_bin_rot_bin_3_count_mean": 50.666666666666664, "action_bin_rot_bin_3_count_std": 2.6246692913372702, "action_bin_rot_bin_3_drop_rate_mean": 0.013502454991816695, "action_bin_rot_bin_3_drop_rate_std": 0.009584138620164138, "action_bin_rot_bin_3_prob_delta_mean_mean": 0.317360336354724, "action_bin_rot_bin_3_prob_delta_mean_std": 0.041824028277672405, "action_bin_rot_bin_3_success_delta_mean_mean": 0.20112919330101184, "action_bin_rot_bin_3_success_delta_mean_std": 0.07564670350551023, "action_bin_trans_bin_0_count_mean": 28.333333333333332, "action_bin_trans_bin_0_count_std": 4.109609335312651, "action_bin_trans_bin_0_drop_rate_mean": 0.10387230627110687, "action_bin_trans_bin_0_drop_rate_std": 0.013988003792639207, "action_bin_trans_bin_0_prob_delta_mean_mean": -0.04340490132560359, "action_bin_trans_bin_0_prob_delta_mean_std": 0.020344127625399628, "action_bin_trans_bin_0_success_delta_mean_mean": -0.10387230627110687, "action_bin_trans_bin_0_success_delta_mean_std": 0.013988003792639207, "action_bin_trans_bin_1_count_mean": 247.33333333333334, "action_bin_trans_bin_1_count_std": 3.0912061651652345, "action_bin_trans_bin_1_drop_rate_mean": 0.14152101409754245, "action_bin_trans_bin_1_drop_rate_std": 0.006560566710555922, "action_bin_trans_bin_1_prob_delta_mean_mean": -0.03691636539016463, "action_bin_trans_bin_1_prob_delta_mean_std": 0.0061656353988546325, "action_bin_trans_bin_1_success_delta_mean_mean": -0.09302355099409985, "action_bin_trans_bin_1_success_delta_mean_std": 0.012013245371742667, "action_bin_trans_bin_2_count_mean": 630.3333333333334, "action_bin_trans_bin_2_count_std": 12.391753530294071, "action_bin_trans_bin_2_drop_rate_mean": 0.10993730382942603, "action_bin_trans_bin_2_drop_rate_std": 0.007395776907489164, "action_bin_trans_bin_2_prob_delta_mean_mean": 0.15572978814811894, "action_bin_trans_bin_2_prob_delta_mean_std": 0.0052222528710755235, "action_bin_trans_bin_2_success_delta_mean_mean": 0.10483876397821297, "action_bin_trans_bin_2_success_delta_mean_std": 0.036173036156861266, "action_bin_trans_bin_3_count_mean": 394.0, "action_bin_trans_bin_3_count_std": 6.377042156569663, "action_bin_trans_bin_3_drop_rate_mean": 0.07025288060199744, "action_bin_trans_bin_3_drop_rate_std": 0.004716990359537022, "action_bin_trans_bin_3_prob_delta_mean_mean": 0.3574832518596545, "action_bin_trans_bin_3_prob_delta_mean_std": 0.011800797236479066, "action_bin_trans_bin_3_success_delta_mean_mean": 0.23186562081163622, "action_bin_trans_bin_3_success_delta_mean_std": 0.04112814415330276, "action_dim_0_mean_mean": 0.8789223723820387, "action_dim_0_mean_std": 0.004072771939731831, "action_dim_0_saturation_rate_mean": 0.7112820512820512, "action_dim_0_saturation_rate_std": 0.01062788809028855, "action_dim_0_std_mean": 0.18688838292142265, "action_dim_0_std_std": 0.003972944789574835, "action_dim_1_mean_mean": -0.024767375160926558, "action_dim_1_mean_std": 0.006189399630468979, "action_dim_1_saturation_rate_mean": 0.07923076923076922, "action_dim_1_saturation_rate_std": 0.004529108136578383, "action_dim_1_std_mean": 0.4077913361414511, "action_dim_1_std_std": 0.0036768074326949535, "action_dim_2_mean_mean": 0.022725954094927055, "action_dim_2_mean_std": 0.012996372386537792, "action_dim_2_saturation_rate_mean": 0.14666666666666667, "action_dim_2_saturation_rate_std": 0.004714045207910323, "action_dim_2_std_mean": 0.5755561411439951, "action_dim_2_std_std": 0.0033732424669339492, "action_dim_3_mean_mean": 0.26748671866298906, "action_dim_3_mean_std": 0.004133040594537617, "action_dim_3_saturation_rate_mean": 0.0038461538461538464, "action_dim_3_saturation_rate_std": 0.0016617283842071437, "action_dim_3_std_mean": 0.24595971185309531, "action_dim_3_std_std": 0.004983630958218022, "action_dim_4_mean_mean": -0.43068150794609655, "action_dim_4_mean_std": 0.0034720296922068983, "action_dim_4_saturation_rate_mean": 0.29538461538461536, "action_dim_4_saturation_rate_std": 0.009504488714282378, "action_dim_4_std_mean": 0.6333207739982044, "action_dim_4_std_std": 0.004657331729517682, "action_dim_5_mean_mean": 0.0492567388374039, "action_dim_5_mean_std": 0.004424656563937332, "action_dim_5_saturation_rate_mean": 0.017948717948717947, "action_dim_5_saturation_rate_std": 0.003223026946147829, "action_dim_5_std_mean": 0.37161652152361896, "action_dim_5_std_std": 0.0036512851660585022, "action_rotation_norm_mean_mean": 0.8862631574720293, "action_rotation_norm_mean_std": 0.00526318163729562, "action_rotation_norm_std_mean": 0.2714854157170743, "action_rotation_norm_std_std": 0.008143184293474822, "action_saturation_rate_mean": 0.20905982905982903, "action_saturation_rate_std": 0.0014853117262378455, "action_translation_norm_mean_mean": 1.0988645859983353, "action_translation_norm_mean_std": 0.00198176954777288, "action_translation_norm_std_mean": 0.3144137024155445, "action_translation_norm_std_std": 0.004641488000798218, "calibrator_after_brier_mean": 0.19262440971280406, "calibrator_after_brier_std": 0.004668170828457833, "calibrator_before_brier_vs_legacy_mean": 0.14875951186836783, "calibrator_before_brier_vs_legacy_std": 0.0040808248937456385, "calibrator_neg_prob_delta_degradation_auc_mean": 0.5657173369612712, "calibrator_neg_prob_delta_degradation_auc_std": 0.009985569643702893, "calibrator_prob_after_auc_mean": 0.7069078260577456, "calibrator_prob_after_auc_std": 0.008529574121041343, "calibrator_prob_before_auc_vs_legacy_mean": 0.8605873375984959, "calibrator_prob_before_auc_vs_legacy_std": 0.00537263516357932, "calibrator_prob_delta_recovery_auc_mean": 0.5619689993758694, "calibrator_prob_delta_recovery_auc_std": 0.030661788118731183, "calibrator_raw_logit_after_auc_mean": 0.7069078260577456, "calibrator_raw_logit_after_auc_std": 0.008529574121041343, "calibrator_raw_logit_before_auc_vs_legacy_mean": 0.8605873375984959, "calibrator_raw_logit_before_auc_vs_legacy_std": 0.00537263516357932, "corr_dim_0_negative_recovery_mean": -0.1018836739456574, "corr_dim_0_negative_recovery_std": 0.013707957156274328, "corr_dim_0_positive_drop_mean": 0.06370311753690037, "corr_dim_0_positive_drop_std": 0.005744019651163355, "corr_dim_0_prob_delta_mean": 0.44704097595370995, "corr_dim_0_prob_delta_std": 0.014470758626389892, "corr_dim_1_negative_recovery_mean": 0.2737140220421528, "corr_dim_1_negative_recovery_std": 0.06579712736196189, "corr_dim_1_positive_drop_mean": -0.0864288972288847, "corr_dim_1_positive_drop_std": 0.03505126281199094, "corr_dim_1_prob_delta_mean": -0.33565686409932516, "corr_dim_1_prob_delta_std": 0.011181792888251205, "corr_dim_2_negative_recovery_mean": 0.057111369436394076, "corr_dim_2_negative_recovery_std": 0.057930799135863874, "corr_dim_2_positive_drop_mean": -0.05890272365471616, "corr_dim_2_positive_drop_std": 0.02951650310856906, "corr_dim_2_prob_delta_mean": -0.08757818160256169, "corr_dim_2_prob_delta_std": 0.04415082707213164, "corr_dim_3_negative_recovery_mean": -0.027629176139935707, "corr_dim_3_negative_recovery_std": 0.009666834155309974, "corr_dim_3_positive_drop_mean": 0.0759345530372027, "corr_dim_3_positive_drop_std": 0.028128605439537012, "corr_dim_3_prob_delta_mean": -0.0750794140690647, "corr_dim_3_prob_delta_std": 0.032873089214565235, "corr_dim_4_negative_recovery_mean": -0.2742593592593205, "corr_dim_4_negative_recovery_std": 0.06722320750835148, "corr_dim_4_positive_drop_mean": 0.05459477077164284, "corr_dim_4_positive_drop_std": 0.032482677891030716, "corr_dim_4_prob_delta_mean": 0.21045718806735114, "corr_dim_4_prob_delta_std": 0.01713396047935431, "corr_dim_5_negative_recovery_mean": 0.3011503478315483, "corr_dim_5_negative_recovery_std": 0.07092126652061831, "corr_dim_5_positive_drop_mean": -0.012847484149825455, "corr_dim_5_positive_drop_std": 0.02449357068098154, "corr_dim_5_prob_delta_mean": -0.23273571143968508, "corr_dim_5_prob_delta_std": 0.012002608545178433, "corr_rotation_norm_prob_delta_mean": 0.13349313574424984, "corr_rotation_norm_prob_delta_std": 0.04257400338293015, "corr_rotation_norm_success_delta_mean": 0.0484817811768212, "corr_rotation_norm_success_delta_std": 0.016680266036954532, "corr_translation_norm_prob_delta_mean": 0.5342521508258403, "corr_translation_norm_prob_delta_std": 0.013518898692674447, "corr_translation_norm_success_delta_mean": 0.22501227467925589, "corr_translation_norm_success_delta_std": 0.017176386227079395, "evaluation_wall_minutes": 47.979174495783326, "experiment_name": "wo-onl-cal", "label": "wo-onl-cal", "macro_success_lift_ci95_high": 0.13102564102564104, "macro_success_lift_ci95_low": 0.07205128205128204, "macro_success_lift_mean": 0.10102564102564103, "macro_success_lift_std": 0.007142763219068771, "neg_hold_ci95_high": 0.6426828959798976, "neg_hold_ci95_low": 0.5397720906660348, "neg_hold_mean": 0.5920216862507405, "neg_hold_std": 0.006574727432415718, "pos_drop_ci95_high": 0.20415925746595856, "pos_drop_ci95_low": 0.15600755290934193, "pos_drop_mean": 0.1802449722027022, "pos_drop_std": 0.003049678038518123, "prob_delta_mean_ci95_high": 0.19514767861670082, "prob_delta_mean_ci95_low": 0.15885240058043656, "prob_delta_mean_mean": 0.17591223537712355, "prob_delta_mean_std": 0.0055473732013499874, "source_summary_csv": "/rl-grasp-refine/outputs/unseen_test_formal_paper_seed7_episode/wo-onl-cal/summary.csv", "t_cover_delta_ci95_high": -0.005860154238280588, "t_cover_delta_ci95_low": -0.027213432000494968, "t_cover_delta_mean": -0.017471287380660495, "t_cover_delta_std": 0.0018772114095656084, "t_edge_delta_ci95_high": -0.040408466158616825, "t_edge_delta_ci95_low": -0.07188255960750953, "t_edge_delta_mean": -0.05709123373839061, "t_edge_delta_std": 0.004946145433523996 }, { "across_object_lift_iqr_mean": 0.10666666666666665, "across_object_lift_iqr_std": 0.03399346342395188, "across_object_lift_std_mean": 0.0787985234456519, "across_object_lift_std_std": 0.008465255630142896, "action_bin_rot_bin_0_count_mean": 208.0, "action_bin_rot_bin_0_count_std": 2.160246899469287, "action_bin_rot_bin_0_drop_rate_mean": 0.08958278919768009, "action_bin_rot_bin_0_drop_rate_std": 0.015502885556702876, "action_bin_rot_bin_0_prob_delta_mean_mean": 0.047941364799732045, "action_bin_rot_bin_0_prob_delta_mean_std": 0.0049374519333494045, "action_bin_rot_bin_0_success_delta_mean_mean": 0.07084636943430268, "action_bin_rot_bin_0_success_delta_mean_std": 0.032687417015058326, "action_bin_rot_bin_1_count_mean": 785.3333333333334, "action_bin_rot_bin_1_count_std": 10.338708279513883, "action_bin_rot_bin_1_drop_rate_mean": 0.09470715625813265, "action_bin_rot_bin_1_drop_rate_std": 0.0069825987037768775, "action_bin_rot_bin_1_prob_delta_mean_mean": 0.042681015912407444, "action_bin_rot_bin_1_prob_delta_mean_std": 0.0014497344553095614, "action_bin_rot_bin_1_success_delta_mean_mean": 0.08148392496075162, "action_bin_rot_bin_1_success_delta_mean_std": 0.005239988454080386, "action_bin_rot_bin_2_count_mean": 306.3333333333333, "action_bin_rot_bin_2_count_std": 11.728408057172787, "action_bin_rot_bin_2_drop_rate_mean": 0.06512652231677578, "action_bin_rot_bin_2_drop_rate_std": 0.004919058998666089, "action_bin_rot_bin_2_prob_delta_mean_mean": 0.14337093419982153, "action_bin_rot_bin_2_prob_delta_mean_std": 0.010042916390070593, "action_bin_rot_bin_2_success_delta_mean_mean": 0.18355170368386944, "action_bin_rot_bin_2_success_delta_mean_std": 0.023315299153119005, "action_bin_rot_bin_3_count_mean": 0.3333333333333333, "action_bin_rot_bin_3_count_std": 0.4714045207910317, "action_bin_rot_bin_3_drop_rate_mean": 0.0, "action_bin_rot_bin_3_drop_rate_std": 0.0, "action_bin_rot_bin_3_prob_delta_mean_mean": 0.19065201643379187, "action_bin_rot_bin_3_prob_delta_mean_std": 0.0, "action_bin_rot_bin_3_success_delta_mean_mean": 0.0, "action_bin_rot_bin_3_success_delta_mean_std": 0.0, "action_bin_trans_bin_0_count_mean": 190.0, "action_bin_trans_bin_0_count_std": 11.575836902790225, "action_bin_trans_bin_0_drop_rate_mean": 0.07448346244023032, "action_bin_trans_bin_0_drop_rate_std": 0.013303578963822829, "action_bin_trans_bin_0_prob_delta_mean_mean": -0.023642514655378366, "action_bin_trans_bin_0_prob_delta_mean_std": 0.004598219865710195, "action_bin_trans_bin_0_success_delta_mean_mean": -0.046201353972841364, "action_bin_trans_bin_0_success_delta_mean_std": 0.01183450798519193, "action_bin_trans_bin_1_count_mean": 475.6666666666667, "action_bin_trans_bin_1_count_std": 11.897712198383164, "action_bin_trans_bin_1_drop_rate_mean": 0.10242967202898474, "action_bin_trans_bin_1_drop_rate_std": 0.004885129165186917, "action_bin_trans_bin_1_prob_delta_mean_mean": 0.024511197899548958, "action_bin_trans_bin_1_prob_delta_mean_std": 0.0021757878384603183, "action_bin_trans_bin_1_success_delta_mean_mean": 0.015957643103775645, "action_bin_trans_bin_1_success_delta_mean_std": 0.006484342301582661, "action_bin_trans_bin_2_count_mean": 583.6666666666666, "action_bin_trans_bin_2_count_std": 8.993825042154695, "action_bin_trans_bin_2_drop_rate_mean": 0.08175410031521042, "action_bin_trans_bin_2_drop_rate_std": 0.00675478670276399, "action_bin_trans_bin_2_prob_delta_mean_mean": 0.12336259047568725, "action_bin_trans_bin_2_prob_delta_mean_std": 0.003877746091894829, "action_bin_trans_bin_2_success_delta_mean_mean": 0.20489706346258985, "action_bin_trans_bin_2_success_delta_mean_std": 0.011156831322680636, "action_bin_trans_bin_3_count_mean": 50.666666666666664, "action_bin_trans_bin_3_count_std": 3.6817870057290873, "action_bin_trans_bin_3_drop_rate_mean": 0.05228241494226149, "action_bin_trans_bin_3_drop_rate_std": 0.006465824404372549, "action_bin_trans_bin_3_prob_delta_mean_mean": 0.16359842454810858, "action_bin_trans_bin_3_prob_delta_mean_std": 0.02227721891545379, "action_bin_trans_bin_3_success_delta_mean_mean": 0.3163011186029089, "action_bin_trans_bin_3_success_delta_mean_std": 0.007173646368726307, "action_dim_0_mean_mean": 0.6701238159993917, "action_dim_0_mean_std": 0.005987321245375175, "action_dim_0_saturation_rate_mean": 0.2856410256410256, "action_dim_0_saturation_rate_std": 0.005699002761767405, "action_dim_0_std_mean": 0.2785523688697636, "action_dim_0_std_std": 0.005365347551589722, "action_dim_1_mean_mean": 0.016271049199731814, "action_dim_1_mean_std": 0.0015362764872203186, "action_dim_1_saturation_rate_mean": 0.0, "action_dim_1_saturation_rate_std": 0.0, "action_dim_1_std_mean": 0.15346236008462172, "action_dim_1_std_std": 0.0007509818609135055, "action_dim_2_mean_mean": -0.06938086475154612, "action_dim_2_mean_std": 0.003336944516242996, "action_dim_2_saturation_rate_mean": 0.020769230769230772, "action_dim_2_saturation_rate_std": 0.0016617283842071439, "action_dim_2_std_mean": 0.4513216684340313, "action_dim_2_std_std": 0.0010706950200360986, "action_dim_3_mean_mean": 0.0762529960982815, "action_dim_3_mean_std": 0.005135067163292765, "action_dim_3_saturation_rate_mean": 0.0, "action_dim_3_saturation_rate_std": 0.0, "action_dim_3_std_mean": 0.2930874716883294, "action_dim_3_std_std": 0.00211495174319666, "action_dim_4_mean_mean": -0.56479045444598, "action_dim_4_mean_std": 0.0020996157058017113, "action_dim_4_saturation_rate_mean": 0.09256410256410257, "action_dim_4_saturation_rate_std": 0.007597708917963622, "action_dim_4_std_mean": 0.20835268930642262, "action_dim_4_std_std": 0.0016599229840076706, "action_dim_5_mean_mean": -0.07381019279324952, "action_dim_5_mean_std": 0.0024055918133794115, "action_dim_5_saturation_rate_mean": 0.0, "action_dim_5_saturation_rate_std": 0.0, "action_dim_5_std_mean": 0.15955184551303017, "action_dim_5_std_std": 0.003075673040772075, "action_rotation_norm_mean_mean": 0.6585356329024474, "action_rotation_norm_mean_std": 0.0014925873720070646, "action_rotation_norm_std_mean": 0.22671420439984255, "action_rotation_norm_std_std": 0.0010659374665897186, "action_saturation_rate_mean": 0.0664957264957265, "action_saturation_rate_std": 0.001899667587255796, "action_translation_norm_mean_mean": 0.813332740987415, "action_translation_norm_mean_std": 0.003606459380801324, "action_translation_norm_std_mean": 0.312287125166321, "action_translation_norm_std_std": 0.0033443667351234978, "calibrator_after_brier_mean": 0.18351241402151872, "calibrator_after_brier_std": 0.0017890312923307468, "calibrator_before_brier_vs_legacy_mean": 0.17230423081596177, "calibrator_before_brier_vs_legacy_std": 0.0008988379551043211, "calibrator_neg_prob_delta_degradation_auc_mean": 0.5448052959771127, "calibrator_neg_prob_delta_degradation_auc_std": 0.01413307947752749, "calibrator_prob_after_auc_mean": 0.7448076712558452, "calibrator_prob_after_auc_std": 0.005929454682794715, "calibrator_prob_before_auc_vs_legacy_mean": 0.8636692779369394, "calibrator_prob_before_auc_vs_legacy_std": 0.0033594428665192568, "calibrator_prob_delta_recovery_auc_mean": 0.5379678599373039, "calibrator_prob_delta_recovery_auc_std": 0.029306468168556362, "calibrator_raw_logit_after_auc_mean": 0.7448076712558452, "calibrator_raw_logit_after_auc_std": 0.005929454682794715, "calibrator_raw_logit_before_auc_vs_legacy_mean": 0.8636692779369394, "calibrator_raw_logit_before_auc_vs_legacy_std": 0.0033594428665192568, "corr_dim_0_negative_recovery_mean": -0.13285744554684642, "corr_dim_0_negative_recovery_std": 0.05592450379506431, "corr_dim_0_positive_drop_mean": 0.12602098366465067, "corr_dim_0_positive_drop_std": 0.02279597515416872, "corr_dim_0_prob_delta_mean": 0.532726687507402, "corr_dim_0_prob_delta_std": 0.014377822958212523, "corr_dim_1_negative_recovery_mean": -0.13012944835187443, "corr_dim_1_negative_recovery_std": 0.05279201284497617, "corr_dim_1_positive_drop_mean": -0.050814847501052425, "corr_dim_1_positive_drop_std": 0.017448838312792512, "corr_dim_1_prob_delta_mean": 0.3694176339850015, "corr_dim_1_prob_delta_std": 0.006931770416211789, "corr_dim_2_negative_recovery_mean": -0.014274232537267992, "corr_dim_2_negative_recovery_std": 0.038533253484901615, "corr_dim_2_positive_drop_mean": -0.12172915763832044, "corr_dim_2_positive_drop_std": 3.6098617489021956e-05, "corr_dim_2_prob_delta_mean": 0.022020276878066902, "corr_dim_2_prob_delta_std": 0.020770551409938987, "corr_dim_3_negative_recovery_mean": 0.1780225976146568, "corr_dim_3_negative_recovery_std": 0.05558546061440917, "corr_dim_3_positive_drop_mean": 0.001595942108649013, "corr_dim_3_positive_drop_std": 0.0217668716329482, "corr_dim_3_prob_delta_mean": -0.49249750090090744, "corr_dim_3_prob_delta_std": 0.007665941933206334, "corr_dim_4_negative_recovery_mean": 0.20662084880615983, "corr_dim_4_negative_recovery_std": 0.0431299597637104, "corr_dim_4_positive_drop_mean": -0.11925915240539238, "corr_dim_4_positive_drop_std": 0.019433456570471574, "corr_dim_4_prob_delta_mean": -0.46354595913574115, "corr_dim_4_prob_delta_std": 0.02369285558443355, "corr_dim_5_negative_recovery_mean": -0.1985519329000154, "corr_dim_5_negative_recovery_std": 0.040610249259019904, "corr_dim_5_positive_drop_mean": 0.047548852432759, "corr_dim_5_positive_drop_std": 0.013664507245368948, "corr_dim_5_prob_delta_mean": 0.4351688762644888, "corr_dim_5_prob_delta_std": 0.014355229190546274, "corr_rotation_norm_prob_delta_mean": 0.37129968438623145, "corr_rotation_norm_prob_delta_std": 0.02234964411515984, "corr_rotation_norm_success_delta_mean": 0.10707808677172133, "corr_rotation_norm_success_delta_std": 0.019764964166976016, "corr_translation_norm_prob_delta_mean": 0.5410825243221318, "corr_translation_norm_prob_delta_std": 0.012375247732433198, "corr_translation_norm_success_delta_mean": 0.21290931766920593, "corr_translation_norm_success_delta_std": 0.02525561891955553, "evaluation_wall_minutes": 46.29419090496376, "experiment_name": "wo-stb-rwd", "label": "wo-stb-rwd", "macro_success_lift_ci95_high": 0.1292371794871794, "macro_success_lift_ci95_low": 0.07974358974358974, "macro_success_lift_mean": 0.10358974358974359, "macro_success_lift_std": 0.002968163308407763, "neg_hold_ci95_high": 0.6238240479425149, "neg_hold_ci95_low": 0.5082373698374928, "neg_hold_mean": 0.5671671475135774, "neg_hold_std": 0.007172592786607987, "pos_drop_ci95_high": 0.17966828323542813, "pos_drop_ci95_low": 0.12890024281629656, "pos_drop_mean": 0.15336418647453992, "pos_drop_std": 0.0021923367252857573, "prob_delta_mean_ci95_high": 0.07359867003262349, "prob_delta_mean_ci95_low": 0.06136644371817845, "prob_delta_mean_mean": 0.06732549386964481, "prob_delta_mean_std": 0.0033964230600683526, "source_summary_csv": "/rl-grasp-refine/outputs/unseen_test_formal_paper_seed7_episode/wo-stb-rwd/summary.csv", "t_cover_delta_ci95_high": -0.024677983829893037, "t_cover_delta_ci95_low": -0.03624486455254542, "t_cover_delta_mean": -0.03074642589623998, "t_cover_delta_std": 0.002177934234768834, "t_edge_delta_ci95_high": -0.04402493043397539, "t_edge_delta_ci95_low": -0.07108803726150655, "t_edge_delta_mean": -0.057859812806336536, "t_edge_delta_std": 0.005046283221218752 }, { "across_object_lift_iqr_mean": 0.13666666666666663, "across_object_lift_iqr_std": 0.020548046676563257, "across_object_lift_std_mean": 0.08448256577765866, "across_object_lift_std_std": 0.009889618844318888, "action_bin_rot_bin_0_count_mean": 386.6666666666667, "action_bin_rot_bin_0_count_std": 7.717224601860151, "action_bin_rot_bin_0_drop_rate_mean": 0.11428286915110546, "action_bin_rot_bin_0_drop_rate_std": 0.018814375575008164, "action_bin_rot_bin_0_prob_delta_mean_mean": 0.026549113349846016, "action_bin_rot_bin_0_prob_delta_mean_std": 0.0027205198874957334, "action_bin_rot_bin_0_success_delta_mean_mean": 0.00921819657104245, "action_bin_rot_bin_0_success_delta_mean_std": 0.031167099976084373, "action_bin_rot_bin_1_count_mean": 603.0, "action_bin_rot_bin_1_count_std": 16.57307052620807, "action_bin_rot_bin_1_drop_rate_mean": 0.10615935200643047, "action_bin_rot_bin_1_drop_rate_std": 0.004528127254854441, "action_bin_rot_bin_1_prob_delta_mean_mean": 0.06120836045509467, "action_bin_rot_bin_1_prob_delta_mean_std": 0.004417687158480671, "action_bin_rot_bin_1_success_delta_mean_mean": 0.08606134630842477, "action_bin_rot_bin_1_success_delta_mean_std": 0.015800472929648535, "action_bin_rot_bin_2_count_mean": 250.33333333333334, "action_bin_rot_bin_2_count_std": 17.98765008430939, "action_bin_rot_bin_2_drop_rate_mean": 0.06679197089086485, "action_bin_rot_bin_2_drop_rate_std": 0.01593275806378622, "action_bin_rot_bin_2_prob_delta_mean_mean": 0.141448421333103, "action_bin_rot_bin_2_prob_delta_mean_std": 0.005937440124599832, "action_bin_rot_bin_2_success_delta_mean_mean": 0.2438237065953396, "action_bin_rot_bin_2_success_delta_mean_std": 0.046766030397543366, "action_bin_rot_bin_3_count_mean": 60.0, "action_bin_rot_bin_3_count_std": 0.0, "action_bin_rot_bin_3_drop_rate_mean": 0.011111111111111112, "action_bin_rot_bin_3_drop_rate_std": 0.007856742013183862, "action_bin_rot_bin_3_prob_delta_mean_mean": 0.22899923949981707, "action_bin_rot_bin_3_prob_delta_mean_std": 0.010544436823339843, "action_bin_rot_bin_3_success_delta_mean_mean": 0.2333333333333333, "action_bin_rot_bin_3_success_delta_mean_std": 0.036004114991154766, "action_bin_trans_bin_0_count_mean": 22.333333333333332, "action_bin_trans_bin_0_count_std": 3.0912061651652345, "action_bin_trans_bin_0_drop_rate_mean": 0.10944444444444444, "action_bin_trans_bin_0_drop_rate_std": 0.05157399095677855, "action_bin_trans_bin_0_prob_delta_mean_mean": 0.0015513369298725984, "action_bin_trans_bin_0_prob_delta_mean_std": 0.0055846149461941325, "action_bin_trans_bin_0_success_delta_mean_mean": -0.04111111111111111, "action_bin_trans_bin_0_success_delta_mean_std": 0.102065096632577, "action_bin_trans_bin_1_count_mean": 426.3333333333333, "action_bin_trans_bin_1_count_std": 14.197026292697903, "action_bin_trans_bin_1_drop_rate_mean": 0.11204229506654072, "action_bin_trans_bin_1_drop_rate_std": 0.008473832675803955, "action_bin_trans_bin_1_prob_delta_mean_mean": 0.009017628175538485, "action_bin_trans_bin_1_prob_delta_mean_std": 0.005134984844332894, "action_bin_trans_bin_1_success_delta_mean_mean": -0.03914996061001989, "action_bin_trans_bin_1_success_delta_mean_std": 0.011196306236129832, "action_bin_trans_bin_2_count_mean": 653.3333333333334, "action_bin_trans_bin_2_count_std": 26.849374087469688, "action_bin_trans_bin_2_drop_rate_mean": 0.09758331771606109, "action_bin_trans_bin_2_drop_rate_std": 0.012175294246249116, "action_bin_trans_bin_2_prob_delta_mean_mean": 0.09194359565479054, "action_bin_trans_bin_2_prob_delta_mean_std": 0.005235252047111064, "action_bin_trans_bin_2_success_delta_mean_mean": 0.14775315770890993, "action_bin_trans_bin_2_success_delta_mean_std": 0.03553095588871224, "action_bin_trans_bin_3_count_mean": 198.0, "action_bin_trans_bin_3_count_std": 12.328828005937952, "action_bin_trans_bin_3_drop_rate_mean": 0.05971212133776713, "action_bin_trans_bin_3_drop_rate_std": 0.01570467508124249, "action_bin_trans_bin_3_prob_delta_mean_mean": 0.16432745543680483, "action_bin_trans_bin_3_prob_delta_mean_std": 0.009431427747161027, "action_bin_trans_bin_3_success_delta_mean_mean": 0.2591840881009047, "action_bin_trans_bin_3_success_delta_mean_std": 0.026492300278799272, "action_dim_0_mean_mean": 0.8108104596383726, "action_dim_0_mean_std": 0.005696276500105467, "action_dim_0_saturation_rate_mean": 0.5094871794871795, "action_dim_0_saturation_rate_std": 0.0058357982957646356, "action_dim_0_std_mean": 0.21501400746358915, "action_dim_0_std_std": 0.005354359986889605, "action_dim_1_mean_mean": -0.061452624861550974, "action_dim_1_mean_std": 0.0030533407834968124, "action_dim_1_saturation_rate_mean": 0.008974358974358975, "action_dim_1_saturation_rate_std": 0.0026148818018424543, "action_dim_1_std_mean": 0.3270066543134061, "action_dim_1_std_std": 0.005949544324179131, "action_dim_2_mean_mean": 0.021347360153442237, "action_dim_2_mean_std": 0.008448071258351649, "action_dim_2_saturation_rate_mean": 0.05128205128205129, "action_dim_2_saturation_rate_std": 0.008363849451435942, "action_dim_2_std_mean": 0.47746169765285235, "action_dim_2_std_std": 0.004483991850306495, "action_dim_3_mean_mean": -0.1805725536290452, "action_dim_3_mean_std": 0.0022117132934546205, "action_dim_3_saturation_rate_mean": 0.018205128205128204, "action_dim_3_saturation_rate_std": 0.002614881801842454, "action_dim_3_std_mean": 0.4212828497978116, "action_dim_3_std_std": 0.0018514876293248804, "action_dim_4_mean_mean": -0.16615658468459374, "action_dim_4_mean_std": 0.009156885875374643, "action_dim_4_saturation_rate_mean": 0.045641025641025644, "action_dim_4_saturation_rate_std": 0.004275213333470018, "action_dim_4_std_mean": 0.3469833436406748, "action_dim_4_std_std": 0.002577301125917398, "action_dim_5_mean_mean": 0.3182577594307114, "action_dim_5_mean_std": 0.0023886991330460293, "action_dim_5_saturation_rate_mean": 0.0, "action_dim_5_saturation_rate_std": 0.0, "action_dim_5_std_mean": 0.2508155719146868, "action_dim_5_std_std": 0.004955567645042441, "action_rotation_norm_mean_mean": 0.6489813428556387, "action_rotation_norm_mean_std": 0.004241211072175042, "action_rotation_norm_std_mean": 0.3181197855051368, "action_rotation_norm_std_std": 0.004657863161701433, "action_saturation_rate_mean": 0.10559829059829061, "action_saturation_rate_std": 0.0012619506889088368, "action_translation_norm_mean_mean": 0.9777590112979805, "action_translation_norm_mean_std": 0.0019799375551326623, "action_translation_norm_std_mean": 0.2948381626616449, "action_translation_norm_std_std": 0.005563339801514857, "calibrator_after_brier_mean": 0.19093172316834642, "calibrator_after_brier_std": 0.0032934944345778673, "calibrator_before_brier_vs_legacy_mean": 0.17492662761721367, "calibrator_before_brier_vs_legacy_std": 0.0018002736884332952, "calibrator_neg_prob_delta_degradation_auc_mean": 0.578598351001654, "calibrator_neg_prob_delta_degradation_auc_std": 0.0255501575891329, "calibrator_prob_after_auc_mean": 0.7198714719240052, "calibrator_prob_after_auc_std": 0.014972786880994134, "calibrator_prob_before_auc_vs_legacy_mean": 0.8652524979892452, "calibrator_prob_before_auc_vs_legacy_std": 0.006031668861947402, "calibrator_prob_delta_recovery_auc_mean": 0.5038707257242835, "calibrator_prob_delta_recovery_auc_std": 0.006575808597589892, "calibrator_raw_logit_after_auc_mean": 0.7198714719240052, "calibrator_raw_logit_after_auc_std": 0.014972786880994134, "calibrator_raw_logit_before_auc_vs_legacy_mean": 0.8652524979892452, "calibrator_raw_logit_before_auc_vs_legacy_std": 0.006031668861947402, "corr_dim_0_negative_recovery_mean": -0.04950445965674977, "corr_dim_0_negative_recovery_std": 0.05356136321849882, "corr_dim_0_positive_drop_mean": 0.07905494624801486, "corr_dim_0_positive_drop_std": 0.020923516149888077, "corr_dim_0_prob_delta_mean": 0.47148327367014103, "corr_dim_0_prob_delta_std": 0.012391139636295753, "corr_dim_1_negative_recovery_mean": -0.12231886459457737, "corr_dim_1_negative_recovery_std": 0.010839905147935482, "corr_dim_1_positive_drop_mean": -0.11641024879239709, "corr_dim_1_positive_drop_std": 0.06810752639412675, "corr_dim_1_prob_delta_mean": -0.1460498779215085, "corr_dim_1_prob_delta_std": 0.018666284734156203, "corr_dim_2_negative_recovery_mean": 0.0415078987815942, "corr_dim_2_negative_recovery_std": 0.0157443727486369, "corr_dim_2_positive_drop_mean": -0.056695954145886025, "corr_dim_2_positive_drop_std": 0.0385740421245486, "corr_dim_2_prob_delta_mean": -0.01362791109924382, "corr_dim_2_prob_delta_std": 0.013810599651778381, "corr_dim_3_negative_recovery_mean": 0.03858363163703026, "corr_dim_3_negative_recovery_std": 0.03905586638692399, "corr_dim_3_positive_drop_mean": 0.09035006692548496, "corr_dim_3_positive_drop_std": 0.029004694302576597, "corr_dim_3_prob_delta_mean": -0.20556472687162017, "corr_dim_3_prob_delta_std": 0.025149941648150514, "corr_dim_4_negative_recovery_mean": 0.24734469968886316, "corr_dim_4_negative_recovery_std": 0.024994010224698666, "corr_dim_4_positive_drop_mean": 0.030563208904680467, "corr_dim_4_positive_drop_std": 0.015279865288984895, "corr_dim_4_prob_delta_mean": -0.026277612722621375, "corr_dim_4_prob_delta_std": 0.022715344555926817, "corr_dim_5_negative_recovery_mean": -0.15861094260204436, "corr_dim_5_negative_recovery_std": 0.06568304327869713, "corr_dim_5_positive_drop_mean": 0.08433800278510732, "corr_dim_5_positive_drop_std": 0.04013837245920052, "corr_dim_5_prob_delta_mean": 0.5470878177687836, "corr_dim_5_prob_delta_std": 0.005596570406081303, "corr_rotation_norm_prob_delta_mean": 0.4656413335481484, "corr_rotation_norm_prob_delta_std": 0.015032012701824278, "corr_rotation_norm_success_delta_mean": 0.16260577006627944, "corr_rotation_norm_success_delta_std": 0.03716869546446292, "corr_translation_norm_prob_delta_mean": 0.5074020604211341, "corr_translation_norm_prob_delta_std": 0.011949625129389138, "corr_translation_norm_success_delta_mean": 0.20950219028092606, "corr_translation_norm_success_delta_std": 0.028011894415270522, "evaluation_wall_minutes": 45.377441812252314, "experiment_name": "wo-tac-rwd", "label": "wo-tac-rwd", "macro_success_lift_ci95_high": 0.12666666666666668, "macro_success_lift_ci95_low": 0.07358974358974359, "macro_success_lift_mean": 0.09948717948717949, "macro_success_lift_std": 0.015208392437268637, "neg_hold_ci95_high": 0.6284506231661526, "neg_hold_ci95_low": 0.5147055563130609, "neg_hold_mean": 0.5720496857008787, "neg_hold_std": 0.005786602020828212, "pos_drop_ci95_high": 0.19712539644716764, "pos_drop_ci95_low": 0.1487717504951508, "pos_drop_mean": 0.17248415269306264, "pos_drop_std": 0.012615794029016067, "prob_delta_mean_ci95_high": 0.08168237095234936, "prob_delta_mean_ci95_low": 0.0674735880715342, "prob_delta_mean_mean": 0.0741112461070111, "prob_delta_mean_std": 0.003908054035255986, "source_summary_csv": "/rl-grasp-refine/outputs/unseen_test_formal_paper_seed7_episode/wo-tac-rwd/summary.csv", "t_cover_delta_ci95_high": -0.013496774098885691, "t_cover_delta_ci95_low": -0.03057928071468356, "t_cover_delta_mean": -0.02270146052692559, "t_cover_delta_std": 0.005581005560811138, "t_edge_delta_ci95_high": -0.03457976622629511, "t_edge_delta_ci95_low": -0.06358622114320855, "t_edge_delta_mean": -0.04943018935204376, "t_edge_delta_std": 0.007735779439937581 }, { "across_object_lift_iqr_mean": 0.13000000000000003, "across_object_lift_iqr_std": 0.029439202887759478, "across_object_lift_std_mean": 0.08660121130157034, "across_object_lift_std_std": 0.013160870216224379, "action_bin_rot_bin_0_count_mean": 205.0, "action_bin_rot_bin_0_count_std": 17.146428199482248, "action_bin_rot_bin_0_drop_rate_mean": 0.10542331457289893, "action_bin_rot_bin_0_drop_rate_std": 0.008063546905227812, "action_bin_rot_bin_0_prob_delta_mean_mean": 0.07711303651118288, "action_bin_rot_bin_0_prob_delta_mean_std": 0.003163963134129699, "action_bin_rot_bin_0_success_delta_mean_mean": 0.0648241854701296, "action_bin_rot_bin_0_success_delta_mean_std": 0.016852533368785515, "action_bin_rot_bin_1_count_mean": 836.6666666666666, "action_bin_rot_bin_1_count_std": 20.8539897594894, "action_bin_rot_bin_1_drop_rate_mean": 0.09795072362623958, "action_bin_rot_bin_1_drop_rate_std": 0.004356414350908351, "action_bin_rot_bin_1_prob_delta_mean_mean": 0.08322355562179658, "action_bin_rot_bin_1_prob_delta_mean_std": 0.004606548821930191, "action_bin_rot_bin_1_success_delta_mean_mean": 0.10193785863819116, "action_bin_rot_bin_1_success_delta_mean_std": 0.01769359934987651, "action_bin_rot_bin_2_count_mean": 257.3333333333333, "action_bin_rot_bin_2_count_std": 7.408703590297623, "action_bin_rot_bin_2_drop_rate_mean": 0.11528652061429238, "action_bin_rot_bin_2_drop_rate_std": 0.015360432397247013, "action_bin_rot_bin_2_prob_delta_mean_mean": 0.09848203341329044, "action_bin_rot_bin_2_prob_delta_mean_std": 0.003485827433160462, "action_bin_rot_bin_2_success_delta_mean_mean": 0.06640184332836965, "action_bin_rot_bin_2_success_delta_mean_std": 0.021507598085169926, "action_bin_rot_bin_3_count_mean": 1.0, "action_bin_rot_bin_3_count_std": 0.816496580927726, "action_bin_rot_bin_3_drop_rate_mean": 0.0, "action_bin_rot_bin_3_drop_rate_std": 0.0, "action_bin_rot_bin_3_prob_delta_mean_mean": 0.234140986116917, "action_bin_rot_bin_3_prob_delta_mean_std": 0.14724968628813076, "action_bin_rot_bin_3_success_delta_mean_mean": 0.25, "action_bin_rot_bin_3_success_delta_mean_std": 0.25, "action_bin_trans_bin_0_count_mean": 6.333333333333333, "action_bin_trans_bin_0_count_std": 1.699673171197595, "action_bin_trans_bin_0_drop_rate_mean": 0.047619047619047616, "action_bin_trans_bin_0_drop_rate_std": 0.06734350297014738, "action_bin_trans_bin_0_prob_delta_mean_mean": -0.045047363357687144, "action_bin_trans_bin_0_prob_delta_mean_std": 0.03075354533963089, "action_bin_trans_bin_0_success_delta_mean_mean": -0.047619047619047616, "action_bin_trans_bin_0_success_delta_mean_std": 0.06734350297014738, "action_bin_trans_bin_1_count_mean": 278.3333333333333, "action_bin_trans_bin_1_count_std": 5.557777333511022, "action_bin_trans_bin_1_drop_rate_mean": 0.1292656618743575, "action_bin_trans_bin_1_drop_rate_std": 0.0071005331980010635, "action_bin_trans_bin_1_prob_delta_mean_mean": 0.004894802604463382, "action_bin_trans_bin_1_prob_delta_mean_std": 0.0015194396650874053, "action_bin_trans_bin_1_success_delta_mean_mean": -0.0660281747238269, "action_bin_trans_bin_1_success_delta_mean_std": 0.00858381646373093, "action_bin_trans_bin_2_count_mean": 862.3333333333334, "action_bin_trans_bin_2_count_std": 9.177266598624136, "action_bin_trans_bin_2_drop_rate_mean": 0.10738170253517963, "action_bin_trans_bin_2_drop_rate_std": 0.008026774778901926, "action_bin_trans_bin_2_prob_delta_mean_mean": 0.09240531053236352, "action_bin_trans_bin_2_prob_delta_mean_std": 0.0023430508897931365, "action_bin_trans_bin_2_success_delta_mean_mean": 0.10017313485292297, "action_bin_trans_bin_2_success_delta_mean_std": 0.012641937701974776, "action_bin_trans_bin_3_count_mean": 153.0, "action_bin_trans_bin_3_count_std": 3.559026084010437, "action_bin_trans_bin_3_drop_rate_mean": 0.028206991110216916, "action_bin_trans_bin_3_drop_rate_std": 0.007743430251692475, "action_bin_trans_bin_3_prob_delta_mean_mean": 0.19696529973820165, "action_bin_trans_bin_3_prob_delta_mean_std": 0.007980066609583465, "action_bin_trans_bin_3_success_delta_mean_mean": 0.3120526233429459, "action_bin_trans_bin_3_success_delta_mean_std": 0.024981312735258, "action_dim_0_mean_mean": 0.8494108827742868, "action_dim_0_mean_std": 0.0028132480352275682, "action_dim_0_saturation_rate_mean": 0.5482051282051282, "action_dim_0_saturation_rate_std": 0.005267343226711303, "action_dim_0_std_mean": 0.18290700233709853, "action_dim_0_std_std": 0.006350843945473766, "action_dim_1_mean_mean": 0.2576183890954902, "action_dim_1_mean_std": 0.005240818800424523, "action_dim_1_saturation_rate_mean": 0.0, "action_dim_1_saturation_rate_std": 0.0, "action_dim_1_std_mean": 0.16049930597386639, "action_dim_1_std_std": 0.00302280859699122, "action_dim_2_mean_mean": 0.17795718075041153, "action_dim_2_mean_std": 0.005377167224317644, "action_dim_2_saturation_rate_mean": 0.06153846153846154, "action_dim_2_saturation_rate_std": 0.003922322702763678, "action_dim_2_std_mean": 0.4882234861073056, "action_dim_2_std_std": 0.005131732691242268, "action_dim_3_mean_mean": -0.1551106507067855, "action_dim_3_mean_std": 0.005036034732129407, "action_dim_3_saturation_rate_mean": 0.0, "action_dim_3_saturation_rate_std": 0.0, "action_dim_3_std_mean": 0.23316074091869912, "action_dim_3_std_std": 0.004049075740205097, "action_dim_4_mean_mean": -0.4096712488256013, "action_dim_4_mean_std": 0.00263499711142823, "action_dim_4_saturation_rate_mean": 0.00358974358974359, "action_dim_4_saturation_rate_std": 0.00145047544858779, "action_dim_4_std_mean": 0.37932747673083916, "action_dim_4_std_std": 0.002394754927112896, "action_dim_5_mean_mean": 0.09683443753644287, "action_dim_5_mean_std": 0.006896269568514755, "action_dim_5_saturation_rate_mean": 0.0007692307692307691, "action_dim_5_saturation_rate_std": 0.0006280742930213276, "action_dim_5_std_mean": 0.3215108196823769, "action_dim_5_std_std": 0.001423600156991895, "action_rotation_norm_mean_mean": 0.6746755791559781, "action_rotation_norm_mean_std": 0.004612969328881896, "action_rotation_norm_std_mean": 0.2185443550534419, "action_rotation_norm_std_std": 0.004816162669242376, "action_saturation_rate_mean": 0.10235042735042736, "action_saturation_rate_std": 0.0011765726408492492, "action_translation_norm_mean_mean": 1.03172170627613, "action_translation_norm_mean_std": 0.0028728559712219237, "action_translation_norm_std_mean": 0.22975964421806291, "action_translation_norm_std_std": 0.0021643926703604197, "calibrator_after_brier_mean": 0.18236854068721983, "calibrator_after_brier_std": 0.0024918499845190886, "calibrator_before_brier_vs_legacy_mean": 0.16719208046286413, "calibrator_before_brier_vs_legacy_std": 0.0011113119020566356, "calibrator_neg_prob_delta_degradation_auc_mean": 0.5983928892004641, "calibrator_neg_prob_delta_degradation_auc_std": 0.012866451061926739, "calibrator_prob_after_auc_mean": 0.7405659239753041, "calibrator_prob_after_auc_std": 0.012632566482388037, "calibrator_prob_before_auc_vs_legacy_mean": 0.8623390086544708, "calibrator_prob_before_auc_vs_legacy_std": 0.00425077573625485, "calibrator_prob_delta_recovery_auc_mean": 0.5401414643236738, "calibrator_prob_delta_recovery_auc_std": 0.02731740546547282, "calibrator_raw_logit_after_auc_mean": 0.7405659239753041, "calibrator_raw_logit_after_auc_std": 0.012632566482388037, "calibrator_raw_logit_before_auc_vs_legacy_mean": 0.8623390086544708, "calibrator_raw_logit_before_auc_vs_legacy_std": 0.00425077573625485, "corr_dim_0_negative_recovery_mean": -0.08840422398383581, "corr_dim_0_negative_recovery_std": 0.03733837099521156, "corr_dim_0_positive_drop_mean": 0.042558276078244466, "corr_dim_0_positive_drop_std": 0.009988259434733183, "corr_dim_0_prob_delta_mean": 0.38428157851468625, "corr_dim_0_prob_delta_std": 0.022334359126431523, "corr_dim_1_negative_recovery_mean": -0.1967018348306914, "corr_dim_1_negative_recovery_std": 0.02186253347248346, "corr_dim_1_positive_drop_mean": 0.059089647815277986, "corr_dim_1_positive_drop_std": 0.013287185271395227, "corr_dim_1_prob_delta_mean": 0.385267743228279, "corr_dim_1_prob_delta_std": 0.01683200626210914, "corr_dim_2_negative_recovery_mean": -0.03992421361144541, "corr_dim_2_negative_recovery_std": 0.016497761546748122, "corr_dim_2_positive_drop_mean": -0.00890350921236176, "corr_dim_2_positive_drop_std": 0.018901721848166048, "corr_dim_2_prob_delta_mean": 0.15554751878050757, "corr_dim_2_prob_delta_std": 0.027443844627248335, "corr_dim_3_negative_recovery_mean": -0.24902433020710177, "corr_dim_3_negative_recovery_std": 0.008829802239358562, "corr_dim_3_positive_drop_mean": -0.0175859569911634, "corr_dim_3_positive_drop_std": 0.02967945156593964, "corr_dim_3_prob_delta_mean": 0.3598308212826668, "corr_dim_3_prob_delta_std": 0.016527795717144446, "corr_dim_4_negative_recovery_mean": -0.17511639989574804, "corr_dim_4_negative_recovery_std": 0.01359776603062686, "corr_dim_4_positive_drop_mean": 0.0034318155876927395, "corr_dim_4_positive_drop_std": 0.011933927315809787, "corr_dim_4_prob_delta_mean": -0.08822455381207811, "corr_dim_4_prob_delta_std": 0.023402134512393938, "corr_dim_5_negative_recovery_mean": 0.18672427400741298, "corr_dim_5_negative_recovery_std": 0.03456341975253265, "corr_dim_5_positive_drop_mean": 0.041962614070683586, "corr_dim_5_positive_drop_std": 0.013815524087160795, "corr_dim_5_prob_delta_mean": -0.1727971814630611, "corr_dim_5_prob_delta_std": 0.02001620338065807, "corr_rotation_norm_prob_delta_mean": 0.06963862522551041, "corr_rotation_norm_prob_delta_std": 0.01608162638447089, "corr_rotation_norm_success_delta_mean": 0.016559935060105568, "corr_rotation_norm_success_delta_std": 0.011965001245980472, "corr_translation_norm_prob_delta_mean": 0.48013051671737816, "corr_translation_norm_prob_delta_std": 0.013184259883243028, "corr_translation_norm_success_delta_mean": 0.21103550958757678, "corr_translation_norm_success_delta_std": 0.017597628922591778, "evaluation_wall_minutes": 45.623874968701664, "experiment_name": "wo-tac-sem-rwd", "label": "wo-tac-sem-rwd", "macro_success_lift_ci95_high": 0.11666666666666665, "macro_success_lift_ci95_low": 0.06230769230769229, "macro_success_lift_mean": 0.0887179487179487, "macro_success_lift_std": 0.007493148305744842, "neg_hold_ci95_high": 0.6246390243205527, "neg_hold_ci95_low": 0.5117800572283095, "neg_hold_mean": 0.5682747745647158, "neg_hold_std": 0.011489970497878083, "pos_drop_ci95_high": 0.20572730351374277, "pos_drop_ci95_low": 0.1538228381387296, "pos_drop_mean": 0.17934832438174986, "pos_drop_std": 0.005817757074088684, "prob_delta_mean_ci95_high": 0.09417300084210527, "prob_delta_mean_ci95_low": 0.07718802768886514, "prob_delta_mean_mean": 0.08527541113834454, "prob_delta_mean_std": 0.002670944959520799, "source_summary_csv": "/rl-grasp-refine/outputs/unseen_test_formal_paper_seed7_episode/wo-tac-sem-rwd/summary.csv", "t_cover_delta_ci95_high": -0.01630261865908392, "t_cover_delta_ci95_low": -0.03505960782737865, "t_cover_delta_mean": -0.02650838499759513, "t_cover_delta_std": 0.0038402470165563554, "t_edge_delta_ci95_high": -0.04637793213584035, "t_edge_delta_ci95_low": -0.07501795186790061, "t_edge_delta_mean": -0.061308455462895865, "t_edge_delta_std": 0.006011346265252827 } ], "source_root": "/rl-grasp-refine/outputs/unseen_test_formal_paper_seed7_episode" }