pi05-organize-tools-trossen-ai

pi05-organize-tools-trossen-ai

A π₀.₅ (pi0.5) policy fine-tuned with openpi (Physical Intelligence's JAX trainer) for the organize tools task on the Trossen AI Stationary bimanual platform.

Framework: openpi (JAX/XLA) — this is not a LeRobot/PyTorch checkpoint. Load it with the openpi runtime, not lerobot.

Inputs

• Dataset: TrossenRoboticsCommunity/trossen_ai_stationary_organize_tools — LeRobot v2 format (openpi requires v2; v3 is rejected).
• Base model: π₀.₅ (pi0.5) pretrained checkpoint from Physical Intelligence, LoRA fine-tuned.
• Training code: TrossenRobotics/openpi
• Hardware: Trossen AI Stationary (bimanual), trained locally on an RTX 5090.
• Action/state space: 14-dim (per assets/trossen/norm_stats.json).

Configuration

• Steps: 100,000
• Batch size: 8
• Fine-tuning: LoRA
• Norm stats: assets/trossen/norm_stats.json (auto-computed at train time).

How it was trained

Fine-tuned locally on an RTX 5090 using the openpi JAX trainer (LoRA, batch size 8, 100K steps). openpi's first-class π implementation generally outperforms LeRobot's PyTorch π policies. Exported as the final orbax checkpoint (params/ = weights, train_state/ = optimizer state for resuming).

How to use

git clone https://github.com/TrossenRobotics/openpi
# follow openpi serving/inference instructions; point the policy at this repo's params/
# recommended inference rate: 50 Hz

Outputs

• Best checkpoint: step 100,000
• Checkpoint exported: 2025-11-05

Evaluation

<!-- Upload eval clips to videos/ in this repo; this player renders them automatically. -->

<video controls width="640" src="https://huggingface.co/TrossenRoboticsCommunity/pi05-organize-tools-trossen-ai/resolve/main/videos/eval.mp4"></video>

Versions

• r1 (2025-11-05) — initial openpi checkpoint export.