Pi05_few_shot_libero_16

PI0.5 Four-Suite Few-Shot Model for LIBERO — Action Chunk 16

This package contains the ten RLinf-native, consolidated PI0.5 checkpoints from few-shot SFT 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 complete four-suite evaluation results. Results are stored at task granularity: each suite has its own task_result.json with the 10 tasks, 50 episodes per task, success counts, and success rates.

Package contents

Pi05_few_shot_libero_16/
├── global_step_<STEP>/
│   ├── actor/model_state_dict/full_weights.pt
│   ├── physical-intelligence/libero/norm_stats.json
│   ├── eval/summary.json
│   ├── eval/summary.md
│   ├── eval/<SUITE>/task_result.json
│   └── rlinf_model_config.yaml
├── physical-intelligence/libero/norm_stats.json
├── metadata/
│   ├── eval_summary.csv/json
│   ├── task_success_rates.csv/json
│   ├── evaluation_protocol.json
│   ├── fewshot_manifest.json
│   ├── norm_stats_manifest.json
│   └── training_config.yaml
├── model_manifest.json
├── rlinf_model_config.yaml
├── SHA256SUMS
├── LICENSE
└── README.md

Each full_weights.pt is a complete native RLinf model state dict. DCP shards, optimizer state, and scheduler state are intentionally excluded; this package is for inference/evaluation or model initialization, not exact training resume.

Training and data

• Training seed: 42
• Few-shot selection: 10 episodes per task, 400 episodes total, 66,481 frames
• Suites: LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-10
• Action dimension: 7
• Action chunk: 16
• Action horizon: 16
• Denoising steps: 10
• Training config: metadata/training_config.yaml
• Selection manifest: metadata/fewshot_manifest.json

Action convention

The seven action dimensions are six robot motion dimensions followed by the normalized gripper dimension. The gripper convention in the training data, normalization statistics, checkpoint outputs, and evaluation adapter is:

-1 = close, +1 = open

The original source labels (0 = close, 1 = open) were converted before normalization and training. Use the included physical-intelligence/libero/norm_stats.json; do not substitute statistics from the original PI0.5 checkpoint or from the old Pi05_few_shot_libero package.

Evaluation

All four suites and all ten checkpoints completed with training/evaluation seed 42, 50 episodes per task, and 500 trajectories per suite. The exact protocol is in metadata/evaluation_protocol.json; the exact per-task records are under each checkpoint's eval/<suite>/task_result.json.

Success-once overview:

• Aggregate is the equal-weighted result over the four suites (2,000 trajectories per checkpoint). The CSV/JSON indexes in metadata/ and each suite's task result preserve the exact measured values.

Integrity check

From the package root, verify all copied model weights with:

sha256sum -c SHA256SUMS

Loading with RLinf

Set PI05_FEWSHOT_CHECKPOINT_DIR to one global_step_* directory and merge the model block from rlinf_model_config.yaml into the rollout/evaluation config:

export PI05_FEWSHOT_CHECKPOINT_DIR=/absolute/path/to/Pi05_few_shot_libero_16/global_step_16000

The included config sets num_action_chunks: 16, action_chunk: 16, action_horizon: 16, and uses the package-local norm_stats.json.

License and attribution

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