{ "step": 1000, "metrics": { "eval_rew_align/success_auprc_amburger66_robotsmith_rbm_task02_robotsmith": 0.0, "eval_rew_align/positive_success_acc_amburger66_robotsmith_rbm_task02_robotsmith": 0.0, "eval_rew_align/negative_success_acc_amburger66_robotsmith_rbm_task02_robotsmith": 1.0, "eval_rew_align/loss_amburger66_robotsmith_rbm_task02_robotsmith": 3.971992015838623, "eval_rew_align/pearson_amburger66_robotsmith_rbm_task02_robotsmith": 0.9432370135101211, "eval_p_rank/kendall_last_amburger66_robotsmith_rbm_task02_robotsmith": 0.7066666666666668, "eval_p_rank/kendall_rewind_last_amburger66_robotsmith_rbm_task02_robotsmith": 1.0, "eval_p_rank/avg_succ_subopt_diff_last_amburger66_robotsmith_rbm_task02_robotsmith": 0.09039011001586911, "eval_p_rank/min_succ_subopt_diff_last_amburger66_robotsmith_rbm_task02_robotsmith": 0.09039011001586911, "eval_p_rank/max_succ_subopt_diff_last_amburger66_robotsmith_rbm_task02_robotsmith": 0.09039011001586911, "eval_p_rank/avg_subopt_fail_diff_last_amburger66_robotsmith_rbm_task02_robotsmith": 0.1155370354652405, "eval_p_rank/min_subopt_fail_diff_last_amburger66_robotsmith_rbm_task02_robotsmith": 0.1155370354652405, "eval_p_rank/max_subopt_fail_diff_last_amburger66_robotsmith_rbm_task02_robotsmith": 0.1155370354652405, "eval_p_rank/avg_succ_fail_diff_last_amburger66_robotsmith_rbm_task02_robotsmith": 0.2059271454811096, "eval_p_rank/min_succ_fail_diff_last_amburger66_robotsmith_rbm_task02_robotsmith": 0.2059271454811096, "eval_p_rank/max_succ_fail_diff_last_amburger66_robotsmith_rbm_task02_robotsmith": 0.2059271454811096, "eval_p_rank/ranking_acc_last_amburger66_robotsmith_rbm_task02_robotsmith": 0.8533333333333334, "eval_p_rank/ranking_acc_all_pairs_last_amburger66_robotsmith_rbm_task02_robotsmith": 0.8533333333333334, "eval_p_rank/ranking_acc_failure_vs_suboptimal_last_amburger66_robotsmith_rbm_task02_robotsmith": 0.88, "eval_p_rank/ranking_acc_suboptimal_vs_successful_last_amburger66_robotsmith_rbm_task02_robotsmith": 0.8, "eval_p_rank/ranking_acc_failure_vs_successful_last_amburger66_robotsmith_rbm_task02_robotsmith": 0.88, "eval_p_rank/kendall_avg_amburger66_robotsmith_rbm_task02_robotsmith": 0.7066666666666668, "eval_p_rank/kendall_rewind_avg_amburger66_robotsmith_rbm_task02_robotsmith": 1.0, "eval_p_rank/avg_succ_subopt_diff_avg_amburger66_robotsmith_rbm_task02_robotsmith": 0.09039011001586911, "eval_p_rank/min_succ_subopt_diff_avg_amburger66_robotsmith_rbm_task02_robotsmith": 0.09039011001586911, "eval_p_rank/max_succ_subopt_diff_avg_amburger66_robotsmith_rbm_task02_robotsmith": 0.09039011001586911, "eval_p_rank/avg_subopt_fail_diff_avg_amburger66_robotsmith_rbm_task02_robotsmith": 0.1155370354652405, "eval_p_rank/min_subopt_fail_diff_avg_amburger66_robotsmith_rbm_task02_robotsmith": 0.1155370354652405, "eval_p_rank/max_subopt_fail_diff_avg_amburger66_robotsmith_rbm_task02_robotsmith": 0.1155370354652405, "eval_p_rank/avg_succ_fail_diff_avg_amburger66_robotsmith_rbm_task02_robotsmith": 0.2059271454811096, "eval_p_rank/min_succ_fail_diff_avg_amburger66_robotsmith_rbm_task02_robotsmith": 0.2059271454811096, "eval_p_rank/max_succ_fail_diff_avg_amburger66_robotsmith_rbm_task02_robotsmith": 0.2059271454811096, "eval_p_rank/ranking_acc_avg_amburger66_robotsmith_rbm_task02_robotsmith": 0.8533333333333334, "eval_p_rank/ranking_acc_all_pairs_avg_amburger66_robotsmith_rbm_task02_robotsmith": 0.8533333333333334, "eval_p_rank/ranking_acc_failure_vs_suboptimal_avg_amburger66_robotsmith_rbm_task02_robotsmith": 0.88, "eval_p_rank/ranking_acc_suboptimal_vs_successful_avg_amburger66_robotsmith_rbm_task02_robotsmith": 0.8, "eval_p_rank/ranking_acc_failure_vs_successful_avg_amburger66_robotsmith_rbm_task02_robotsmith": 0.88, "eval_p_rank/kendall_sum_amburger66_robotsmith_rbm_task02_robotsmith": 0.7066666666666668, "eval_p_rank/kendall_rewind_sum_amburger66_robotsmith_rbm_task02_robotsmith": 1.0, "eval_p_rank/avg_succ_subopt_diff_sum_amburger66_robotsmith_rbm_task02_robotsmith": 0.09039011001586911, "eval_p_rank/min_succ_subopt_diff_sum_amburger66_robotsmith_rbm_task02_robotsmith": 0.09039011001586911, "eval_p_rank/max_succ_subopt_diff_sum_amburger66_robotsmith_rbm_task02_robotsmith": 0.09039011001586911, "eval_p_rank/avg_subopt_fail_diff_sum_amburger66_robotsmith_rbm_task02_robotsmith": 0.1155370354652405, "eval_p_rank/min_subopt_fail_diff_sum_amburger66_robotsmith_rbm_task02_robotsmith": 0.1155370354652405, "eval_p_rank/max_subopt_fail_diff_sum_amburger66_robotsmith_rbm_task02_robotsmith": 0.1155370354652405, "eval_p_rank/avg_succ_fail_diff_sum_amburger66_robotsmith_rbm_task02_robotsmith": 0.2059271454811096, "eval_p_rank/min_succ_fail_diff_sum_amburger66_robotsmith_rbm_task02_robotsmith": 0.2059271454811096, "eval_p_rank/max_succ_fail_diff_sum_amburger66_robotsmith_rbm_task02_robotsmith": 0.2059271454811096, "eval_p_rank/ranking_acc_sum_amburger66_robotsmith_rbm_task02_robotsmith": 0.8533333333333334, "eval_p_rank/ranking_acc_all_pairs_sum_amburger66_robotsmith_rbm_task02_robotsmith": 0.8533333333333334, "eval_p_rank/ranking_acc_failure_vs_suboptimal_sum_amburger66_robotsmith_rbm_task02_robotsmith": 0.88, "eval_p_rank/ranking_acc_suboptimal_vs_successful_sum_amburger66_robotsmith_rbm_task02_robotsmith": 0.8, "eval_p_rank/ranking_acc_failure_vs_successful_sum_amburger66_robotsmith_rbm_task02_robotsmith": 0.88, "time/custom_evaluations": 16.21657526679337 } }