π0.5 fine-tuned from pi05_droid on the FR3 bussing task.

Action space: 7 joint velocities (rad/s) + gripper position (8-d). Run at 5 Hz (min_period: 0.2) — the training tick and execution tick are both 0.2 s. flip_gripper: True — trained on the DROID gripper convention (1-g of the robot's raw values); the runner flips state in and action out.

The previous revision baked a ×3 velocity factor into the norm stats after training — each executed tick covered ~3 frames of demo path while the model's chunk clock advanced one. This revision trains the speed-up natively: velocity targets are derived across a 3-frame stride, (q[t+3]-q[t])/0.2, so one 0.2 s tick covers 3 frames of path by construction. Timeline exact, 3× demo speed, no post-hoc scaling anywhere (VELOCITY_SCALE.json = 1.0).

Training data: idle runs cut with episodes split at every cut (no chunk spans a cut), grasp/release events explicitly protected from the idle cutter (without that rescue 107 of 292 gripper transitions sat inside cut regions), 3 phase-offset sub-episodes so every original frame is used.

Register pi05_droid_bussing_5hz (see train_config_snippet.py), norm stats ship in assets/bussing_droid_5hz_s3/. Action horizon 16, first 8 executed. Instruction: put the vegetables in the bowl.