DreamMachines/actuator_unboxing_speedcond_t2_fullft_bs256

pi0.5 speed conditioning, test 2 (real 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 labelled slow, merged with 437 episodes of the same task demonstrated about 1.75× faster in another session, labelled fast (635 episodes, 371,523 frames; 50 fps, three 224×224 cameras, 14-D state and action). 30 % of training samples get unknown, the null prompt for classifier-free guidance.

Training: 4,400 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): the token is not used. Switching it moves the predicted chunk by 0.057, below the sampling noise of 0.063, and the fast/slow ratio of predicted per-step motion is 1.00. The policy predicts the faster motion from the observation instead (1.6–1.8× on frames of the faster session under either prompt): the two sessions are distinguishable from the images, so the token adds no information. The artificial-speed model (actuator_unboxing_speedcond_t1_fullft_bs256), where the images are identical, does learn it.