Hugging Face's logo

THU-SIGS-EILAB
/
Pi05_few_shot_libero

Robotics
rlinf
pi05
vision-language-action
libero
few-shot
Model card Files
xet
 Community
Pi05_few_shot_libero / global_step_8000 /README.md
Monita108's picture
Monita108
Update PI05 few-shot model documentation
91965fa verified
preview code
|
Raw
History Blame Contribute Delete
2.24 kB

PI0.5 few-shot checkpoint: global_step_8000

This directory contains the RLinf-native PI0.5 checkpoint saved at training step 8,000 from the Pi05_few_shot_libero few-shot run. It is an independent, inference-ready checkpoint and includes its own normalization statistics and consolidated LIBERO evaluation results.

Model description

  • Base policy: PI0.5 / OpenPI, config name pi05_libero
  • Training data: 10 selected episodes per task, sampling seed 42
  • Suites represented in the run: LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-10
  • Action dimension: 7 (six robot motion dimensions plus one gripper dimension)
  • Gripper convention: normalized value 0 = close, 1 = open
  • Inference: 10 denoising steps, noise level 0.5, five action chunks per inference

The gripper convention is part of the model/evaluation contract. Keep the seventh action value in [0, 1] and do not reverse its meaning in a custom adapter.

Evaluation result

The checkpoint was evaluated on all four suites with 50 episodes per task, 500 trajectories per suite, fixed ordered reset states, and evaluation seed 195. The aggregate success-once rate is 0.670 (1340/2000).

Suite Trajectories Success-once Success-at-end
LIBERO-Spatial 500 0.782 0.170
LIBERO-Object 500 0.848 0.158
LIBERO-Goal 500 0.658 0.114
LIBERO-10 500 0.392 0.054

The complete reusable parameters are in eval/summary.json, and per-task results are in the suite subdirectories under eval/.

Files and loading

  • actor/model_state_dict/full_weights.pt: complete model state dict
  • physical-intelligence/libero/norm_stats.json: required normalization statistics for this checkpoint
  • eval/summary.json: evaluation parameters and aggregate metrics
  • rlinf_model_config.yaml: portable RLinf model fragment

Set the checkpoint directory before evaluation:

export PI05_FEWSHOT_CHECKPOINT_DIR=/absolute/path/to/Pi05_few_shot_libero/global_step_8000

Then use the included rlinf_model_config.yaml and keep openpi_data.norm_stats_path pointed at the normalization file in this directory.