Samisaliveagain/so100_smolvla_rc

SO-100 SmolVLA Nut-Stacking Dataset

This is a real-robot LeRobot v3 dataset collected with an SO-100/SO-101-style follower arm for the task:

Pick the nut on the right and stack it on the left one.

The dataset combines the original 80-episode v3–v10 collection with the 80 episodes from v13–v20. The v12 recording was intentionally excluded.

Dataset summary

Property	Value
Episodes	160
Frames	77,104
Frame rate	30 FPS
Approximate duration	42.8 minutes
Tasks	1
Robot type	so_follower
LeRobot format	v3.0
Split	train (0:160)

Observations and actions

Each frame contains:

• observation.images.top: 640×480 RGB top-view video, AV1, 30 FPS
• observation.images.fpv: 640×480 RGB first-person-view video, AV1, 30 FPS
• observation.state: six joint positions
• action: six target joint positions
• timestamps and frame, episode, task, and global indices

The six state/action dimensions are:

1. shoulder_pan.pos
2. shoulder_lift.pos
3. elbow_flex.pos
4. wrist_flex.pos
5. wrist_roll.pos
6. gripper.pos

Collection and provenance

Recorded jointly by Samisaliveagain and Shubham. This publication uses the complete 160-episode dataset from local folder 2. The source repository is shubham4413/so100_smolvla_rc.

The dataset was recorded by teleoperating the physical follower arm with a leader arm. It is an aggregate of these sources, in order:

• merged_v3_v10_80ep: 80 episodes
• v13 through v20: 10 episodes each

Episode indices were regenerated during aggregation and are contiguous from 0 through 159. Frame indices are contiguous from 0 through 77,103. The two camera streams and the Parquet metadata were validated after aggregation.

Loading with LeRobot

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("Samisaliveagain/so100_smolvla_rc")
print(dataset.num_episodes, dataset.num_frames)

Intended uses

This dataset is intended for robot-learning research, imitation learning, SmolVLA fine-tuning, offline analysis, and reproducible experiments on the recorded nut-stacking task.

Limitations and safety

• The data represents one robot setup, workspace, camera arrangement, and task.
• It does not establish robustness to new objects, lighting, viewpoints, or hardware calibration.
• The dataset has no held-out evaluation split and should not be used alone to claim real-world policy performance.
• The videos have not been exhaustively audited for incidental sensitive or identifying content.
• Policies trained on this data must be tested with conservative joint limits, low speeds, an emergency stop, and human supervision before physical use.

Related checkpoint

The batch-size-16, 20,000-step SmolVLA checkpoint was trained only on the first 80 episodes (the eight original v3–v10 shards), not on all 160 episodes in this repository.

License

No dataset license has been declared. Contact the dataset owner before reuse or redistribution beyond uses permitted by applicable law.