edge-inference/smolvla-so101-pick-orange-data

SO101 Pick Orange — Teleoperation Dataset

Teleoperated demonstrations of a SO101 robot arm picking three oranges and placing them on a plate in NVIDIA Isaac Sim.

Dataset Details

Property	Value
Robot	SO101 (6-DOF, Feetech STS3215)
Task	Pick 3 oranges → place on plate → return to rest
Environment	NVIDIA Isaac Sim 5.1.0 (kitchen scene)
Teleop Device	SO101 leader arm (ZMQ, 50Hz)
Format	LeRobot v3.0
FPS	30
Episodes	100
Total Frames	60,541

Features

Feature	Shape	Type
action	(6,)	float32 — joint positions [shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper]
observation.state	(6,)	float32 — current joint positions (same order)
observation.images.front	(480, 640, 3)	video (AV1, MP4) — fixed overhead camera
observation.images.wrist	(480, 640, 3)	video (AV1, MP4) — wrist-mounted camera

Domain Randomization

Applied per-episode at reset:

• Orange positions: ±3 cm (XY)
• Plate position: ±6 cm (XY)
• Front camera: ±2.5 cm position, ±2.5° orientation

Collection Method

Demonstrations collected via SO101 leader-follower teleoperation in simulation. Only successful episodes (all 3 oranges placed on plate) are included — failures are discarded during recording.

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("edge-inference/smolvla-so101-pick-orange")
print(f"{dataset.num_episodes} episodes, {dataset.num_frames} frames")

Intended Use

Fine-tuning vision-language-action models (e.g., SmolVLA) for sim-to-real transfer on SO101 pick-and-place tasks.

Authors

• Benjamin Kubwimana (@benyamini)

Part of the Edge Inference project.

Citation

If you use this dataset in your research, please cite:

@misc{kubwimana2026so101pickorange,
  title={SO101 Pick Orange Teleoperation Dataset},
  author={Kubwimana, Benjamin},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/edge-inference/smolvla-so101-pick-orange}
}

Acknowledgements

This dataset was created using leisaac with NVIDIA Isaac Sim and the LeRobot framework.