TRON2 Pick-and-Place PI0.5
Fine-tuned PI0.5 policy for the instruction:
place the yellow box inside the blue bin
Model contract
- Robot state/action dimension: 16
- Action horizon: 50
- Cameras:
cam_high,cam_left_wrist,cam_right_wrist - Actions: absolute joint and gripper commands
- Training-time RTC simulated delay: 10
- Training data: 200 LeRobot v2.1 episodes
- Training exposure: 20,000 steps at global batch size 32
- Training hardware: 4 NVIDIA H200 GPUs
- Source Slurm job: 19566304
The normalization statistics in assets/tron2_pick_place_200/norm_stats.json
are part of the model contract and must be used for inference.
Safety
Validate image ordering, state ordering, action ordering, output shape, finite values, and joint limits in shadow mode before sending actions to a robot.