π0.5 Piper Pick-Bowl-Into-Bag LoRA
OpenPI/JAX π0.5 LoRA checkpoint trained for 8,000 steps on
pick_up_bowl_ep009to040_32_right7d_pi05_rgb224_v21 (32 episodes, 18,129 frames).
The exact language prompt is:
pick up the bowl and place it in the bag
Runtime contract
- RGB inputs: top and right wrist, letterboxed to 224×224.
- State/action: 7-D
right [j1..j6, gripper_m]; the six joint targets are trained as state-relative deltas and the gripper remains absolute.
- Recorded action supervision contains absolute joint-angle targets in radians.
- Grippers are meters and are clipped to
[0, 0.08] at the policy boundary.
- Action horizon: 30.
- OpenPI config:
pi05_piper_pick_bowl_lora.
Normalization statistics under assets/ and the bundled OpenPI overlay are
required. The validation report checks checkpoint loading and inference pipeline
integrity; it is not a real-robot success-rate evaluation.
