pi05-piper-throw-into-tank-replay40-lora

π0.5 Piper Throw-Into-Tank LoRA

OpenPI/JAX LoRA checkpoint trained only on the 40 demonstrations from Throw_Into_Tank_Clean_Background_Replay_Current_Camera (15,658 frames). These action replays were recorded with the current camera setup; episode 032 is the hash-verified corrected teleoperation from Throw01_episode32_fixing. The older Throw_Into_Tank01 and the _2, _3, _4 sessions are not part of this run. No demonstrations were held out from training. The fixed task prompt is throw the object into the tank.

Runtime contract

• Two RGB inputs: top and right wrist, letterboxed to 224×224.
• State/action order: [R_j1..j6, R_gripper] (7-D).
• Arm joints are radians; grippers are meters.
• Recorded actions are absolute joint targets.
• During training, six arm joints are converted to deltas against current state; the gripper remains absolute.
• During inference, arm outputs are converted back to absolute targets and grippers are clipped to [0, 0.08] meters.
• Action horizon is 30 and the returned action vector is 7-D, routed only to the right arm by the robot client.

Use the bundled OpenPI overlay and configuration pi05_piper_throw_tank_replay40_lora. Normalization statistics under assets/ are required. The validation/ directory contains pipeline smoke reports; these test checkpoint loading and inference integrity, not real-robot task success.

Validate with conservative robot-side safety limits before commanding hardware.