VLABench/pi0-base-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.

The reference result of this model is:

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	0.52	0.74	0.6	0.8	0.24	0.24	0.62	0.753	0.326	0.424	0.461
track_2_cross_category	0.24	0.58	0.54	0.77	0.22	0.22	0.2	0.24	0.049	0.098	0.25
track_3_common_sense	0.1	0.49	0	0.18	0.38	0.38	0.2	0.247	0.091	0.125	0.154
track_4_semantic_instruction	0.1	0.47	0	0.18	0.34	0.34	0	0.213	0.021	0.074	0.092
track_6_unseen_texture	0.54	0.76	0.66	0.82	0.18	0.18	0.42	0.647	0.306	0.388	0.421