THU-SIGS-EILAB/Pi05_few_shot_libero

PI0.5 Four-Suite Few-Shot Model for LIBERO

This package contains all ten RLinf-native, consolidated PI0.5 checkpoints from a few-shot fine-tuning run on libero_spatial, libero_object, libero_goal, and libero_10. Checkpoints are provided every 2,000 steps from global_step_2000 through global_step_20000.

Every checkpoint includes a complete four-suite eval/ directory with 500 trajectories per suite. All ten checkpoints have 40 task results and 2,000 trajectories in total.

Package contents

Pi05_few_shot_libero/
├── global_step_2000/
│   ├── actor/model_state_dict/full_weights.pt
│   ├── physical-intelligence/libero/norm_stats.json
│   ├── eval/
│   │   ├── summary.json
│   │   ├── summary.md
│   │   ├── libero_spatial/task_result.json
│   │   ├── libero_object/task_result.json
│   │   ├── libero_goal/task_result.json
│   │   └── libero_10/task_result.json
│   └── rlinf_model_config.yaml
├── global_step_4000/
│   └── ...
├── ...
├── global_step_20000/
│   └── ...
├── physical-intelligence/libero/norm_stats.json
├── metadata/
│   ├── eval_config_libero_10.yaml
│   ├── eval_summary.csv
│   ├── eval_summary.json
│   ├── fewshot_manifest.json
│   ├── norm_stats_manifest.json
│   └── training_config.yaml
├── model_manifest.json
├── rlinf_model_config.yaml
├── SHA256SUMS
├── LICENSE
└── README.md

Every full_weights.pt is a complete model state dict, not a delta or adapter. No base-model weight file is required when loading one with the compatible RLinf PI0.5 implementation. Each global_step_* directory includes its own copy of the required normalization statistics and is directly usable as a model directory. DCP shards and optimizer/scheduler state are intentionally not included, so this package is intended for inference/evaluation or model initialization, not exact training-state resume.

Few-shot data

• Sampling seed: 42
• Sampling mode: 10 episodes per task
• Suites: LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-10
• Selected data: 400 episodes and 66,481 frames
• State dimension: 8
• Action dimension: 7

The exact selected episode IDs and selection fingerprint are recorded in metadata/fewshot_manifest.json. Always use the included few-shot norm_stats.json; substituting the normalization statistics from another PI0.5 checkpoint changes the policy input/output contract.

Evaluation

The common protocol uses 10 tasks per suite, 50 episodes per task, 500 trajectories per suite, evaluation seed 195, fixed and ordered reset states, ten denoising steps, noise level 0.5, and five action chunks per inference.

Success-once overview:

Step	Coverage	Spatial	Object	Goal	LIBERO-10	Aggregate*
2,000	4/4	0.122	0.394	0.208	0.080	0.201
4,000	4/4	0.440	0.610	0.488	0.134	0.418
6,000	4/4	0.694	0.690	0.652	0.396	0.608
8,000	4/4	0.782	0.848	0.658	0.392	0.670
10,000	4/4	0.752	0.890	0.658	0.474	0.694
12,000	4/4	0.776	0.822	0.748	0.488	0.709
14,000	4/4	0.746	0.862	0.774	0.468	0.713
16,000	4/4	0.762	0.922	0.728	0.480	0.723
18,000	4/4	0.702	0.818	0.584	0.522	0.656
20,000	4/4	0.848	0.878	0.742	0.566	0.758

• Aggregate is weighted over all four suites (2,000 trajectories). Each global_step_*/eval/summary.json records the reusable evaluation parameters, and each suite directory contains per-task success counts and rates.

Integrity check

From the package root, verify every checkpoint after transfer with:

sha256sum -c SHA256SUMS

This reads all ten weight files (about 85.3 GB in total), so it can take several minutes on network storage.

Loading with RLinf

Use an RLinf checkout containing the native OpenPI checkpoint loader. Point PI05_FEWSHOT_CHECKPOINT_DIR at the chosen global_step_* directory and merge the model block from rlinf_model_config.yaml into the rollout/evaluation config. For example, to use the formally evaluated checkpoint:

export PI05_FEWSHOT_CHECKPOINT_DIR=/absolute/path/to/Pi05_few_shot_libero/global_step_16000

The important settings are:

rollout:
  model:
    model_type: openpi
    model_path: ${oc.env:PI05_FEWSHOT_CHECKPOINT_DIR}
    openpi:
      config_name: pi05_libero
      checkpoint_format: native
    openpi_data:
      norm_stats_path: ${oc.env:PI05_FEWSHOT_CHECKPOINT_DIR}/physical-intelligence/libero/norm_stats.json

For reproduction, reference training and LIBERO-10 evaluation configs are retained under metadata/; set their environment variables to paths in the new environment before running.

License and attribution

The accompanying RLinf materials are provided under the Apache License 2.0; see LICENSE. Users should also follow the licenses and attribution requirements of PI0.5/OpenPI and LIBERO.