VLABench/pi0-fast-primitive

The following pi0-fast weights were obtained by training on 4 A100 GPUs for 10k iterations using five tasks (2500 episodes) from the primitive-ft-dataset, shared for community reference.

The five primitive tasks used to train are: [select_fruit, select_toy, select_painting, select_poker, select_mahjong]. These tasks involve similar skills and simple actions, making them suitable for research on downstream adaptation and generalization abilities.

The training codes are available at: https://github.com/Shiduo-zh/openpi. If any issues or bugs are encountered during training, feel free to contact our team.

Track	select_toy_SR	select_toy_PS	select_fruit_SR	select_fruit_PS	select_painting_SR	select_painting_PS	select_poker_SR	select_poker_PS	select_mahjong_SR	select_mahjong_PS	Avg_SR
track_1_in_distribution	0.34	0.6	0.44	0.67	0.24	0.24	0.22	0.387	0.25	0.417	0.298
track_2_cross_category	0.16	0.45	0.48	0.67	0.12	0.12	0.14	0.227	0.143	0.204	0.209
track_3_common_sense	0.28	0.54	0.24	0.4	0.18	0.18	0.04	0.193	0.064	0.064	0.161
track_4_semantic_instruction	0.22	0.5	0.294	0.471	0.08	0.08	0.08	0.207	0.146	0.208	0.164
track_6_unseen_texture	0.28	0.55	0.52	0.7	0.24	0.24	0.26	0.367	0.319	0.436	0.324