pi0.5 RoboTwin Policy
This repository contains a LeRobot pi0.5 policy fine-tuned for RoboTwin manipulation tasks.
Model details
- Base model: pi0.5
- Vision-language backbone: PaliGemma / Gemma 2B
- Action expert: Gemma 300M
- Training steps: 50,000
- Training dataset:
official_robotwin_full9_95_5_train - Input: three RGB camera views and a 14-dimensional robot state
- Output: 14-dimensional robot action
- Action chunk size: 50
- Inference steps: 10
Intended use
Use this checkpoint for research evaluation and robot manipulation experiments in RoboTwin with the LeRobot pi0.5 inference pipeline.
Download
hf download <your-username>/<your-repo> --local-dir ./pi05_robotwin
The downloaded directory can be passed to the existing LeRobot pi0.5 policy server as its checkpoint directory.
Evaluation
The model is evaluated on a 9-task RoboTwin suite using qpos actions. The table below merges 30-episode evaluation.
| Task | Success rate |
|---|---|
pick_dual_bottles |
45.2% |
place_dual_shoes |
45.2% |
move_pillbottle_pad |
54.8% |
move_can_pot |
32.3% |
place_mouse_pad |
48.4% |
dump_bin_bigbin |
87.1% |
beat_block_hammer |
93.5% |
rotate_qrcode |
77.4% |
place_a2b_left |
77.4% |
Limitations
This checkpoint is intended for research use. Performance can vary across tasks, seeds, simulator versions, hardware, and camera or state preprocessing settings. It has not been validated for safety-critical or real-world deployment.
License
No separate license is specified for this checkpoint. Please check the licenses of the base model, dataset, and LeRobot before redistribution or commercial use.
- Downloads last month
- 49