smolvla-cucumber-grab-place-peel-doris082126
smolvla policy trained via Vibe Data Studio on 744 teleop episodes.
At a glance
| Architecture | smolvla (base: lerobot/smolvla_base) |
| Dataset | cucumber-place-elevator-home-doris082026-v1-trim @ 90107bd6 |
| Scale | 744 episodes · 95,399 frames @ 20 fps |
| Steps / batch / seed | 40,000 / 32 / 42 |
| lerobot | d114d3b1 |
| Provider / GPU | runpod / NVIDIA A100-SXM4-80GB · pod video preprocess (512px H.264, RGB tracks only) |
| Duration | 4h 53m |
| Final train loss | 0.052 |
| W&B | vds34-913e6ab5 |
Tasks
Language-conditioned — pass one of these instructions at inference:
| Instruction | Episodes |
|---|---|
| place the cucumber on the middle of the cutting board, aligned vertically in the image, with the back end of the cucumber under the blue gripper | 264 |
| grab the cucumber close to one of the cucumber's end | 201 |
| peel the skin of the cucumber with several strokes, starting close to the blue gripper and moving vertically along the cucumber toward the robot origin | 188 |
| place the cucumber on the middle of the cutting board, aligned vertically in the image, with the back end of the cucumber close to the upper edge of the cutting board | 91 |
Episodes per task per recording rig (derived from the nested collection manifests down to leaf recorded datasets):
| Task | jetson1 | unrecorded:cucumber-place-DAgger-iter0-v1-trim | vibepi | Total |
|---|---|---|---|---|
| place the cucumber on the middle of the cutting board, alig… | 150 | 50 | 64 | 264 |
| grab the cucumber close to one of the cucumber's end | 142 | 0 | 59 | 201 |
| peel the skin of the cucumber with several strokes, startin… | 117 | 0 | 71 | 188 |
| place the cucumber on the middle of the cutting board, alig… | 49 | 0 | 42 | 91 |
| Total | 458 | 50 | 236 | 744 |
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.camera1,observation.images.camera2,observation.images.camera3 - Camera renames at training (feed the same mapping at inference):
base→camera1,top→camera2,wrist→camera3
Data provenance
Materialized collection cucumber-place-elevator-home-doris082026@v1 (frozen 2026-08-21):
vibepi-place-elevator-home-doris082026@ba53af98(42 eps)jetson1-place-elevator-home-doris081926@88cfecf6(49 eps)cucumber-grab-place-peel-doris072326-v1-flat@61e62b18(653 eps)
lerobot-train --policy.path=lerobot/smolvla_base --policy.repo_id=VibeCuisine/smolvla-cucumber-grab-place-peel-doris082126 --dataset.repo_id=VibeCuisine/cucumber-place-elevator-home-doris082026-v1-trim --dataset.revision=90107bd63b9f1ca65743c571afa5ece62798fddd --batch_size=32 --steps=40000 --output_dir=outputs/train/smolvla-cucumber-grab-place-peel-doris082126 --job_name=smolvla-cucumber-grab-place-peel-doris082126 --policy.device=cuda --wandb.enable=true --wandb.disable_artifact=true --wandb.project=lerobot --wandb.run_id=vds34-913e6ab5 --wandb.entity=jeremyhx-freelance --seed=42 --rename_map='{"observation.images.base": "observation.images.camera1", "observation.images.top": "observation.images.camera2", "observation.images.wrist": "observation.images.camera3"}' --dataset.eval_split=0.1 --eval_steps=800
--num_workers is appended host-side (clamped to the host's allocation).
Card generated by Vibe Data Studio on run completion.