oku006/bussing-pi05-droid-fix

⚠ SUPERSEDED — do not deploy

The current, tested deliverables are:

• π0.5: oku006/bussing-pi05-droid via branch pi05-joint-velocity of sombit888/robot_evals
• whiteboard: oku006/bussing-whiteboard-act via branch whiteboard-joint-velocity

This repository is kept only as history/provenance. (21 Aug 2026)

---

bussing-pi05-droid — fix_h16, velocity ×3

π0.5 fine-tuned from pi05_droid on the FR3 table-bussing recording. Action space: 7 joint velocities (rad/s) + gripper position (8-d).

This revision replaces the first fine-tune. Three deliberate changes:

1. Gripper is DROID convention (1-g of the robot's raw values), so the pretrained model is not asked to invert what it already knows. Deploy with flip_gripper: True — the runner flips state in and action out.
2. Trained on bussing_droid_fix: images re-decoded from the original mp4s, leading idle trimmed (the operator waited a mean of 7 s before touching the controller; those frames pair a ready scene with a zero action and taught the arm to stand still).
3. Norm stats carry a ×3 velocity factor (VELOCITY_SCALE.json). The demos average 0.03 rad/s — near-invisible on the arm — and normalization is linear, so scaling the action stats is mathematically identical to having trained on 3×-faster demos. Cold-scene gate: 8/10 scenes move and aim right (cosine 0.79 vs the operator's actual opening move), against 0/10 unscaled. Do NOT scale again downstream.

Model: action_horizon=16, execute the first 8, replan. action_dim=32, first 8 dims real. Norm stats asset id: bussing_droid_fix (ships in assets/, loaded from the checkpoint dir — no external download).

Register the config from train_config_snippet.py, then:

policy = create_trained_policy(get_config("pi05_droid_bussing_fix"), CKPT_DIR)

Held-out episodes [2, 4, 10, 20, 21] were never trained on, in any revision.