LGG100/rm-delivery-v1

pi05_rm65_4tasks_350 (v1)

pi0.5 fine-tuned on the RealMan RM65 right arm: pick up the rightmost object and retract while keeping it grasped. 4 objects, 350 episodes (2026-10-05 delivery, merged then trimmed: rm65/4tasks_350_v2_trim, 123,698 frames @ 30 fps).

• openpi config: pi05_rm65_4tasks_350 (LLM LoRA + full action expert, action_horizon 32, EMA 0.99, batch 32, 20k steps)
• state / action: 7-dim absolute joints right_q1..q6 (rad) + right_gripper (0 closed .. 100 open)
• cameras: observation/top, observation/wrist
• checkpoints: 12000/, 16000/, 19999/ — each has params/ and assets/rm65/4tasks_350_v2_trim/norm_stats.json (optimizer train_state/ omitted). Final train loss ≈ 0.0016.

Prompts (use verbatim):

• Pick up the rightmost yellow original-flavor potato chip package as seen in the top camera image with the right arm, then retract the arm while keeping it securely grasped.
• Pick up the rightmost green-and-red cucumber-flavor potato chip package as seen in the top camera image with the right arm, then retract the arm while keeping it securely grasped.
• Pick up the rightmost purple tomato-flavor potato chip package with the right arm, then retract the arm while keeping it securely grasped.
• Pick up the rightmost Coca-Cola container containing black liquid as seen in the top camera image with the right arm, then retract the arm while keeping it securely grasped.

Serve:

uv run scripts/serve_policy.py policy:checkpoint --policy.config=pi05_rm65_4tasks_350 --policy.dir=<path>/19999