actuator_unboxing_speedcond_t1_fullft_bs256

pi0.5 speed conditioning, test 1 (artificial speed)

Full fine-tune of lerobot/pi05_base on a bimanual actuator-unboxing task, with a speed token in the text prompt:

Take an actuator from the box and place it in the tray. Speed: <slow|fast|unknown>

Part of a check whether pi0.5 can be conditioned through text tokens, as a precursor to advantage conditioning (RECAP, π*0.6).

Data (not on the Hub): 198 teleoperated episodes (50 fps, three 224×224 cameras, 14-D state and action) labelled slow, plus two 2× copies of every episode that keep only the even or only the odd frames, labelled fast (594 episodes, 348,616 frames). Slow and fast samples show the same images, so only the token tells them apart. 30 % of training samples get unknown, the null prompt for classifier-free guidance.

Training: 4,000 steps (about 3 epochs), batch 256, learning rate 5e-5, 10 % linear warm-up, cosine decay over the last 25 % to 2.5e-6, full fine-tune on 8×H100, with the Dream Machines fork of LeRobot.

Result on held-out episodes (8 sampled chunks per prompt): switching the token moves the predicted chunk by 0.438, 9.3× the sampling noise of 0.047. Predicted per-step motion under fast is 1.76× that under slow (the data ratio is 1.92), and unknown lies in between.