DreamMachines/actuator_unboxing_20h_345_v2

Actuator unboxing, 21 h (main training set)

The full collection: all 7,271 episodes rated 3 or better, from both operators and all five scenes. This is the export every 50k-step run in the post trains on unless stated otherwise.

One of the nine training datasets behind the Dream Machines post Empirical results fine-tuning π0.5 on a real manufacturing task (September 2026), which fine-tunes π0.5 on this task and reports what mattered: training duration, LoRA vs. full fine-tuning, data quantity, scene diversity, data quality, intervention data and serving settings. Used in Section 2.1 (the 76% baseline), Section 2.2 (the two EBS-64 LoRA arms), Section 2.3, Section 2.4, Section 2.6 (the 21h arm), Section 2.10 and Section 2.11.

The task

Actuators arrive in cardboard boxes and must be transferred into a tray that fixes their position and orientation for a downstream production line. Each actuator is inserted with its bottom contacts correctly oriented, with both possible 180° rotations acceptable. Only the top layer of the source box is used: 40 actuators. The left arm picks an actuator from the box and hands it to the right arm, which inserts it into one of the tray's 135 holes, arranged in 9 rows of 15. The order is fixed: actuators leave the box right to left and bottom to top, and fill the tray left to right and bottom to top. Every episode starts and ends in a rest position.

The post describes the task, the collection and the evaluation protocol in full: Section 1.

This dataset

• The complete filtered collection described below.
• actuator_unboxing_20h_345 holds the same episodes in an earlier export (per-episode video files, 21.4 h vs 21.0 h here).

Collection

• Robot: bimanual TRLC-DK1 (bi_dk1_follower), recorded at 50 Hz.
• Cameras: three fisheye cameras (left wrist, right wrist, top context) recorded at 640×480 MJPG and centre-cropped to 224×224 in this export, matching π0.5's SigLIP input size.
• Teleoperation: two operators, one with a VR headset (our open-source VR teleop kit), one with Gello leader arms.
• Scenes: four office desks and a lightbox, day and night; box and tray positions randomised independently over about 11×11 cm.
• Filtering: every episode was rated 1 to 5 (1 an irrecoverable failure, 3 a minor mistake, 5 a clean execution); only episodes rated 3 or above are kept, and idle periods at the start and end are trimmed.

Contents

Episodes	7,271
Frames	3,782,725 (21.0 h at 50 Hz)
Cameras	3: left_wrist, right_wrist, top, 224×224
State / action	14-dim: six joint positions and one gripper per arm, absolute joint targets
Format	LeRobot v3.0

Results in the post

• Full fine-tune of π0.5, 50k steps at batch 256, the post's baseline: 76% (91/120).
• The same checkpoint with tuned real-time-chunking settings (reinfer 50, blend 20): 93% (74/80).
• Learning rate (2e-5 to 1e-4), batch size (64 to 2048), relative joint targets and image augmentation all landed within the noise of a 40-rollout evaluation.
• LoRA rank 64 at EBS 64: 33% (13/40); rank 256 at EBS 64: 53% (21/40).
• The same recipe from the PaliGemma backbone alone, without π0.5's robot pretraining: 0/40.

All evaluations: 40 rollouts per round in the lightbox, one per actuator of the top layer, end-to-end success with Wilson 95% intervals; see the evaluation protocol.

Related datasets

• DreamMachines/actuator_unboxing_20h_345: Actuator unboxing, 21 h (earlier export)
• DreamMachines/actuator_unboxing_10h_vr: Actuator unboxing, operator 1 (VR teleoperation), 10 h
• DreamMachines/actuator_unboxing_10h_leaderarms: Actuator unboxing, operator 2 (Gello leader arms), 10 h
• DreamMachines/actuator_unboxing_10h_diverse: Actuator unboxing, 10 h across operators and scenes
• DreamMachines/actuator_unboxing_4h_diverse: Actuator unboxing, 4 h across operators and scenes
• DreamMachines/actuator_unboxing_4h_lightbox: Actuator unboxing, 4 h from the lightbox only
• DreamMachines/actuator_unboxing_1h_HQ: Actuator unboxing, one high-quality hour
• DreamMachines/actuator_unboxing_1h_HQ_dagger_intv: Actuator unboxing, one high-quality hour plus intervention data

Citation

@misc{dreammachines2026pi05,
  title  = {Empirical results fine-tuning π0.5 on a real manufacturing task},
  author = {Dream Machines},
  year   = {2026},
  url    = {https://dream-machines.eu/blog/pi05-fine-tuning}
}