hackhackhack66666's picture
upload sim_eval_phase_b/eval_log.json
255c2c9 verified
Raw
History Blame Contribute Delete
4.27 kB
{
"KITCHEN_SCENE3_turn_on_the_stove_and_put_the_moka_pot_on_it/mean_success_rate_mean": 0.8866666666666667,
"KITCHEN_SCENE3_turn_on_the_stove_and_put_the_moka_pot_on_it/mean_success_rate_std": 0.04163331998932269,
"KITCHEN_SCENE3_turn_on_the_stove_and_put_the_moka_pot_on_it/mean_success_rate_stderr": 0.024037008503093284,
"KITCHEN_SCENE4_put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/mean_success_rate_mean": 0.9466666666666667,
"KITCHEN_SCENE4_put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/mean_success_rate_std": 0.011547005383792525,
"KITCHEN_SCENE4_put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/mean_success_rate_stderr": 0.006666666666666672,
"KITCHEN_SCENE6_put_the_yellow_and_white_mug_in_the_microwave_and_close_it/mean_success_rate_mean": 0.7399999999999999,
"KITCHEN_SCENE6_put_the_yellow_and_white_mug_in_the_microwave_and_close_it/mean_success_rate_std": 0.040000000000000036,
"KITCHEN_SCENE6_put_the_yellow_and_white_mug_in_the_microwave_and_close_it/mean_success_rate_stderr": 0.023094010767585053,
"KITCHEN_SCENE8_put_both_moka_pots_on_the_stove/mean_success_rate_mean": 0.43333333333333335,
"KITCHEN_SCENE8_put_both_moka_pots_on_the_stove/mean_success_rate_std": 0.03055050463303893,
"KITCHEN_SCENE8_put_both_moka_pots_on_the_stove/mean_success_rate_stderr": 0.017638342073763937,
"LIVING_ROOM_SCENE1_put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/mean_success_rate_mean": 0.0,
"LIVING_ROOM_SCENE1_put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/mean_success_rate_std": 0.0,
"LIVING_ROOM_SCENE1_put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/mean_success_rate_stderr": 0.0,
"LIVING_ROOM_SCENE2_put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/mean_success_rate_mean": 0.006666666666666667,
"LIVING_ROOM_SCENE2_put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/mean_success_rate_std": 0.011547005383792516,
"LIVING_ROOM_SCENE2_put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/mean_success_rate_stderr": 0.006666666666666667,
"LIVING_ROOM_SCENE2_put_both_the_cream_cheese_box_and_the_butter_in_the_basket/mean_success_rate_mean": 0.5666666666666668,
"LIVING_ROOM_SCENE2_put_both_the_cream_cheese_box_and_the_butter_in_the_basket/mean_success_rate_std": 0.07571877794400365,
"LIVING_ROOM_SCENE2_put_both_the_cream_cheese_box_and_the_butter_in_the_basket/mean_success_rate_stderr": 0.04371625682868001,
"LIVING_ROOM_SCENE5_put_the_white_mug_on_the_left_plate_and_put_the_yellow_and_white_mug_on_the_right_plate/mean_success_rate_mean": 0.48,
"LIVING_ROOM_SCENE5_put_the_white_mug_on_the_left_plate_and_put_the_yellow_and_white_mug_on_the_right_plate/mean_success_rate_std": 0.11135528725660043,
"LIVING_ROOM_SCENE5_put_the_white_mug_on_the_left_plate_and_put_the_yellow_and_white_mug_on_the_right_plate/mean_success_rate_stderr": 0.06429100507328636,
"LIVING_ROOM_SCENE6_put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/mean_success_rate_mean": 0.34,
"LIVING_ROOM_SCENE6_put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/mean_success_rate_std": 0.10583005244258362,
"LIVING_ROOM_SCENE6_put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/mean_success_rate_stderr": 0.061101009266077866,
"STUDY_SCENE1_pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/mean_success_rate_mean": 0.66,
"STUDY_SCENE1_pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/mean_success_rate_std": 0.11135528725660043,
"STUDY_SCENE1_pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/mean_success_rate_stderr": 0.06429100507328636,
"checkpoint": "/workspace/OAT-RoboMimic-Fine-tune/BLT-OAT/output/long/oat_dense_with_uid_long_0530_220204/checkpoints/ep-0950_sr-0.527.ckpt",
"episodes_per_task_approx": 50.0,
"mean_success_rate_mean": 0.506,
"mean_success_rate_std": 0.013114877048604012,
"mean_success_rate_stderr": 0.007571877794400371,
"n_test": 500,
"n_test_per_task": 50,
"num_exp": 3,
"num_tasks": 10,
"seed_stride": 500,
"task_suite": "libero10",
"test_start_seed": 1000
}