{ "step": 250, "metrics": { "eval_rew_align/mse_roboarena_eval": 0.0572185566740965, "eval_rew_align/pearson_roboarena_eval": -0.20521791406560833, "eval_p_rank/spearman_roboarena_eval": 0.1486367653896934, "eval_p_rank/spearman_rewind_roboarena_eval": 0.285297764198399, "eval_p_rank/ranking_acc_roboarena_eval": 0.5525760135135135, "eval_p_rank/ranking_total_pairs_roboarena_eval": 4736.0, "eval_rew_align/mse_mw_eval": 0.028775711768743873, "eval_rew_align/pearson_mw_eval": -0.2622550899752058, "eval_p_rank/spearman_mw_eval": 0.21948339897935923, "eval_p_rank/spearman_rewind_mw_eval": 0.23529411764705882, "eval_p_rank/avg_succ_subopt_diff_mw_eval": 0.00025275735294118495, "eval_p_rank/min_succ_subopt_diff_mw_eval": -0.013867187499999989, "eval_p_rank/max_succ_subopt_diff_mw_eval": 0.017187500000000022, "eval_p_rank/avg_subopt_fail_diff_mw_eval": 0.0027573529411764708, "eval_p_rank/min_subopt_fail_diff_mw_eval": -0.0166015625, "eval_p_rank/max_subopt_fail_diff_mw_eval": 0.0234375, "eval_p_rank/avg_succ_fail_diff_mw_eval": 0.0030101102941176554, "eval_p_rank/min_succ_fail_diff_mw_eval": -0.012890624999999989, "eval_p_rank/max_succ_fail_diff_mw_eval": 0.02128906250000001, "eval_p_rank/ranking_acc_mw_eval": 0.6078431372549019, "eval_p_rank/ranking_total_pairs_mw_eval": 51.0, "eval_p_rank/spearman_utd_so101": -0.02834936490538903, "eval_p_rank/spearman_rewind_utd_so101": -0.1, "eval_p_rank/avg_succ_subopt_diff_utd_so101": -0.00234375, "eval_p_rank/min_succ_subopt_diff_utd_so101": -0.029296875, "eval_p_rank/max_succ_subopt_diff_utd_so101": 0.029296875, "eval_p_rank/avg_subopt_fail_diff_utd_so101": 0.0, "eval_p_rank/min_subopt_fail_diff_utd_so101": -0.0361328125, "eval_p_rank/max_subopt_fail_diff_utd_so101": 0.0341796875, "eval_p_rank/avg_succ_fail_diff_utd_so101": -0.00234375, "eval_p_rank/min_succ_fail_diff_utd_so101": -0.0224609375, "eval_p_rank/max_succ_fail_diff_utd_so101": 0.021484375, "eval_p_rank/ranking_acc_utd_so101": 0.4666666666666667, "eval_p_rank/ranking_total_pairs_utd_so101": 30.0, "eval_p_rank/spearman_usc_franka": 0.359375, "eval_p_rank/spearman_rewind_usc_franka": 0.25, "eval_p_rank/avg_succ_subopt_diff_usc_franka": 0.003173828125, "eval_p_rank/min_succ_subopt_diff_usc_franka": -0.001953125, "eval_p_rank/max_succ_subopt_diff_usc_franka": 0.0107421875, "eval_p_rank/avg_subopt_fail_diff_usc_franka": 0.00244140625, "eval_p_rank/min_subopt_fail_diff_usc_franka": -0.0107421875, "eval_p_rank/max_subopt_fail_diff_usc_franka": 0.017578125, "eval_p_rank/avg_succ_fail_diff_usc_franka": 0.005615234375, "eval_p_rank/min_succ_fail_diff_usc_franka": -0.0126953125, "eval_p_rank/max_succ_fail_diff_usc_franka": 0.0166015625, "eval_p_rank/ranking_acc_usc_franka": 0.5833333333333334, "eval_p_rank/ranking_total_pairs_usc_franka": 12.0, "eval_p_rank/spearman_usc_xarm": 0.05691772515768497, "eval_p_rank/spearman_rewind_usc_xarm": 0.08333333333333333, "eval_p_rank/avg_succ_subopt_diff_usc_xarm": 0.005045572916666667, "eval_p_rank/min_succ_subopt_diff_usc_xarm": -0.015625, "eval_p_rank/max_succ_subopt_diff_usc_xarm": 0.0185546875, "eval_p_rank/avg_subopt_fail_diff_usc_xarm": -0.004069010416666667, "eval_p_rank/min_subopt_fail_diff_usc_xarm": -0.0205078125, "eval_p_rank/max_subopt_fail_diff_usc_xarm": 0.01171875, "eval_p_rank/avg_succ_fail_diff_usc_xarm": 0.0009765625, "eval_p_rank/min_succ_fail_diff_usc_xarm": -0.01171875, "eval_p_rank/max_succ_fail_diff_usc_xarm": 0.013671875, "eval_p_rank/ranking_acc_usc_xarm": 0.5555555555555556, "eval_p_rank/ranking_total_pairs_usc_xarm": 18.0, "time/custom_evaluations": 45.65803809789941 } }