pi0-omy-pnp

PI0 (Physical Intelligence π₀) policy trained on a mujoco pick-and-place task with an omy 6-DoF robot arm.

Task

• Task: Place the red or blue mug on the plate, driven by language instruction
  • "Place the red mug on the plate."
  • "Place the blue mug on the plate."
• Observation: agent camera + wrist camera (256×256) + 6-DoF joint state
• Action: 7-dim (6 joints + gripper)
• Dataset: 20 teleoperated demos (2621 frames, 20fps)

Training

• Policy: PI0 (PaliGemma + action expert), chunk size 50
• Steps: 20,000 (checkpoints at 10k / 20k)
• Hardware: AMD GPU (ROCm 7.x), ~0.83 s/step
• Framework: lerobot (source build, commit 10b7b35), transformers 4.57.6

Closed-loop evaluation (held-out seeds)

4 environment seeds (1000-1003) with re-randomized object placements — all passed:

seed	physical_success	xy_dist (m)	max_lift (m)	stable_place_steps
1000	✅	0.056	0.135	5
1001	✅	0.031	0.118	5
1002	✅	0.035	0.131	5
1003	✅	0.043	0.127	5

Success criterion: mug placed within plate area and stable for ≥5 steps (physical_success), verified by both legacy and physical checks.

Files

• config.json — model configuration
• model.safetensors — trained weights (~7.5 GB)
• train_config.json — training configuration

Usage

from lerobot.common.policies.pi0.modeling_pi0 import PI0Policy

policy = PI0Policy.from_pretrained("howexixi12/pi0-omy-pnp")