vds-4g-utilization-check-20260909

<!-- vibedata:model-card -->

vds-4g-utilization-check-20260909

molmoact2 policy trained via Vibe Data Studio on 188 teleop episodes.

At a glance

Tasks

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

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

Rig

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

Data provenance

Materialized collection cucumber-peel-DAgger-iter2@v2 (frozen 2026-07-22):

• cucumber-peel-DAgger-iter0-v1-trim @ 3604420c (42 eps)
• jetson1-peel-doris072026 @ ba099f15 (47 eps)
• vibepi-peel-doris072126 @ 04d8dcf5 (2 eps)
• vibepi-peel-doris072226 @ 8ed1cb6d (42 eps)
• vibepi-peel-adaptive-doris072226 @ 111b736f (27 eps)
• jetson1-peel-adaptive-doris072026 @ a1ad7824 (28 eps)

<details> <summary><b>Reproduce</b> — exact training command</summary>

lerobot-train --policy.type=molmoact2 --policy.checkpoint_path=allenai/MolmoAct2 --policy.repo_id=VibeCuisine/vds-4g-utilization-check-20260909 --dataset.repo_id=VibeCuisine/cucumber-peel-DAgger-iter2-v2-trim --dataset.revision=692d1ac83fea08c85b1b1bdb1bf25e6551f8cf41 --batch_size=8 --steps=100 --save_freq=2000 --output_dir=outputs/train/vds-4g-utilization-check-20260909 --job_name=vds-4g-utilization-check-20260909 --policy.device=cuda --wandb.enable=true --wandb.disable_artifact=true --wandb.project=lerobot --wandb.run_id=vds61-a10004d5 --wandb.entity=jeremyhx-freelance --seed=42 --dataset.image_transforms.enable=true --policy.dtype=bfloat16 --policy.action_mode=continuous --policy.inference_action_mode=continuous --policy.control_mode='absolute joint pose' --policy.normalize_gripper=true --policy.setup_type='single robotic arm' --policy.gradient_checkpointing=true --policy.train_mode_vlm=lora --policy.enable_inference_cuda_graph=false --policy.chunk_size=30 --policy.n_action_steps=30 --policy.normalization_mapping='{"ACTION":"QUANTILES","STATE":"QUANTILES","VISUAL":"IDENTITY"}' --policy.scheduler_decay_steps=100 --policy.scheduler_warmup_steps=10 --save_freq=100 --log_freq=10

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

</details>

---

Card generated by Vibe Data Studio on run completion.