vibepi3-grab-act-aug-r3

act policy trained via Vibe Data Studio on 64 teleop episodes.

At a glance

Architecture	act (from scratch)
Dataset	vibepi3-grab-poseexpert-r3-curated @ 740f9046
Scale	64 episodes · 3,478 frames @ 20 fps
Steps / batch / seed	30,000 / 8 / 1000
lerobot	b6d46c40
Provider / GPU	site-gpu / site GPU host
Duration	1h 17m
Final train loss	0.047
W&B	vds48-2fc878f0

Tasks

Language-conditioned — pass one of these instructions at inference:

Instruction	Episodes
Grab the cucumber at the one-third point	64

Episodes per task per recording rig (derived from the nested collection manifests down to leaf recorded datasets):

Task	vibepi3	Total
Grab the cucumber at the one-third point	64	64
Total	64	64

Rig

• Action/state (7-dim): shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper, tilt
• Cameras @ 20 fps: corner 640×480, top 640×480, wrist 640×480
• Consumed image inputs (the checkpoint's resolved VISUAL features, from its config.json): observation.images.corner, observation.images.top, observation.images.wrist

Reproduce — exact training command

lerobot-train --policy.type=act --policy.repo_id=VibeCuisine/vibepi3-grab-act-aug-r3 --dataset.repo_id=VibeCuisine/vibepi3-grab-poseexpert-r3-curated --dataset.revision=740f9046b9b9f80c038c27aea6fb35854ac3a708 --batch_size=8 --steps=30000 --save_freq=5000 --output_dir=outputs/train/vibepi3-grab-act-aug-r3 --job_name=vibepi3-grab-act-aug-r3 --policy.device=cuda --wandb.enable=true --wandb.disable_artifact=true --wandb.project=lerobot --wandb.run_id=vds48-2fc878f0 --wandb.entity=jeremyhx-freelance --seed=1000 --dataset.eval_split=0.1 --eval_steps=600 --dataset.image_transforms.enable=true --policy.chunk_size=40 --policy.n_action_steps=40

--num_workers is appended host-side (clamped to the host's allocation).

---

Card generated by Vibe Data Studio on run completion.