Pi0-fast delta action chunk trained on VLABench datasets.
This repository provides the Pi0-fast model trained with the whole VLABench's official primitive tasks dataset. To be noticed, this config corresponds to the delta chunk, which is state[1:] - state[:-1] instead of state - state[0].
Evaluation
To run this checkpoint, please clone this repo: https://github.com/Shiduo-zh/openpi, and checkout to the branch main.
Assume that you download this checkpoints and put it in the directory checkpoints, to run the policy as server, please run:
bash vla_bench_scipts/serve_policy.sh pi0_ft_vlabench_primitive_aligned checkpoints/VLABench/pi0-fast-ft-primitive-10task-deltachunk/29999/
After serving the policy, open another terminal and run:
bash vla_bench_scipts/multi_run_vlabench.sh <Your path to store the evaluate results>
Train
To reproduce the training result, please run the training script with the config pi0_ft_vlabench_primitive_aligned.
XLA_PYTHON_CLIENT_MEM_FRACTION=0.95 uv run scripts/train.py pi0_ft_vlabench_primitive_aligned --exp-name=pi0_ft_vlabench_primitive_aligned --overwrite
Our checkpoint is trained on 8 H100 for 30k iterations, with 5000 episodes data acrossing 10 tasks.
Reference Results
The reference success rate of this model is:
| Track | add_condiment | insert_flower | select_book | select_chemistry_tube | select_drink | select_fruit | select_mahjong | select_painting | select_poker | select_toy | Avg_SR |
|---|---|---|---|---|---|---|---|---|---|---|---|
| track_1_in_distribution | 0.48 | 0.38 | 0.24 | 0.62 | 0.18 | 0.76 | 0.52 | 0.46 | 0.78 | 0.7 | 0.512 |
| track_2_cross_category | 0.02 | ? | 0.04 | 0.14 | 0.06 | 0.74 | 0.25 | 0.46 | 0.72 | 0.44 | 0.342 |
| track_3_common_sense | 0.44 | 0.36 | 0.16 | 0.88 | 0.18 | 0.56 | 0.02 | 0.48 | 0.34 | 0.52 | 0.417 |
| track_4_semantic_instruction | 0.42 | 0.34 | 0.18 | 0.6 | 0.08 | 0.62 | 0.5 | 0.5 | 0.7 | 0.48 | 0.442 |
| track_6_unseen_texture | 0.56 | 0.32 | 0.24 | 0.52 | 0.24 | 0.6 | 0.36 | 0.38 | 0.54 | 0.58 | 0.444 |