Twu31/so101_hand_blue_napkin

SO-ARM101 — "Hand me the blue napkin"

Teleoperated demonstrations of a human–robot handover on a real SO-ARM101 — a 6-DOF, ~$300 open-source arm with STS3215 servos. The robot picks up a pack of blue tissues from the table and places it into a human hand.

Recorded with LeRobot (codebase_version: v2.1).

Robot	so101_follower, 6 DOF
Task	"Hand me the blue napkin" (single task)
Episodes	101 (complete set)
Frames	40,493
Episode length	400–401 frames ≈ 13.4 s each (very consistent)
Frame rate	30 fps
Cameras	2 × 480×640 RGB, AV1-encoded MP4
Total duration	≈ 22.5 minutes
Collection	human teleoperation (leader–follower)

This is the full set the SmolVLA policy was trained on — an earlier release contained only a 43-episode subset.

Trained policies

Twu31/smolvla-so101-blue-napkin-160k	SmolVLA, 160k steps — 80%+ real-arm success (production)
Twu31/smolvla-so101-blue-napkin-50k	earlier 50k-step checkpoint
Twu31/so101_hand_blue_napkin_eval_rollouts	recorded real-arm eval rollouts from both

Features

Key	Type	Shape	Contents
observation.state	float32	(6,)	shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, gripper.pos
action	float32	(6,)	same six joint targets, as commanded by the teleoperator
observation.images.front	video	(480, 640, 3)	fixed front-facing camera
observation.images.handeye	video	(480, 640, 3)	wrist-mounted (hand-eye) camera
timestamp, frame_index, episode_index, index, task_index	—	—	standard LeRobot indexing

Joint positions are in the SO-ARM101 servo's native normalised position units, not radians.

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("Twu31/so101_hand_blue_napkin")
print(ds.num_episodes, ds.num_frames)     # 101 40493

sample = ds[0]
sample["observation.state"]               # (6,)
sample["action"]                          # (6,)
sample["observation.images.front"]        # (3, 480, 640)
sample["observation.images.handeye"]      # (3, 480, 640)
sample["task"]                            # "Hand me the blue napkin"

Or just the tabular part, without decoding video:

from datasets import load_dataset
ds = load_dataset("Twu31/so101_hand_blue_napkin")

Action chunking works well here: at 30 fps a 50-step chunk covers 1.7 s, and episodes are uniform enough that chunk boundaries land in comparable phases of the motion.

What this data has been used for

Two projects, both documented at https://github.com/twu3202/soarm101-offline-rl-experiments:

1. A working SmolVLA policy. A SmolVLA v2 policy (SmolVLM2-500M backbone + flow-matching action expert) was fine-tuned on this dataset and reached 80%+ success on the real arm — 101 demos, 160k steps. The checkpoint is published at Twu31/smolvla-so101-blue-napkin-160k.

2. A negative result on offline RL. Five offline-RL approaches (IQL with several reward labellings, Vision-IQL on frozen SigLIP features, and Residual IQL on top of SmolVLA) were tried on these demonstrations to push past the IL baseline. The best variant improved held-out action MSE by 51% over SmolVLA alone (0.0039 vs 0.0087) and then scored 0% on the real arm — per-frame residuals break SmolVLA's chunk-level self-correction and the closed-loop distribution shift cascades.

   The transferable lesson for anyone using this dataset: open-loop action MSE on held-out demonstrations does not predict closed-loop success. Evaluate in the loop.

Structure

data/chunk-000/episode_0000{00..100}.parquet    # 6D state + 6D action + indices, 101 files
videos/chunk-000/observation.images.front/      # 101 × MP4 (AV1)
videos/chunk-000/observation.images.handeye/    # 101 × MP4 (AV1)
meta/info.json                                  # feature schema, fps, counts
meta/episodes.jsonl                             # per-episode length and task
meta/episodes_stats.jsonl                       # per-episode per-feature min/max/mean/std
meta/tasks.jsonl                                # the single task string

Limitations

• One task, one scene, one operator. Fixed camera placement, fixed table, the same tissue pack throughout. Do not expect a policy trained on this to generalise to other objects, backgrounds, or receiving hands.
• 101 episodes is still small in absolute terms, and there is no held-out split defined — splits declares all 101 as train. If you need a validation set, carve one out yourself.
• Successful demonstrations only. There are no failure episodes, no recovery behaviour, and no reward or success annotations. Anything reward-shaped has to be derived (the project above did this from gripper events and joint-velocity-based success-frame detection).
• AV1 video. Decoding needs a recent av/ffmpeg; older installs will fail to read frames.
• The human hand's position varies between episodes, so the same arm state can map to different correct actions. This multi-modality is what made state-only IQL degenerate to behaviour cloning in the experiments above — worth knowing before fitting a unimodal policy.

Privacy

A human hand and forearm appear in the frames of both cameras. No faces, no identifying features, no audio.

Citation

@dataset{so101_hand_blue_napkin_2026,
  author    = {Twu31},
  title     = {SO-ARM101 — "Hand me the blue napkin": teleoperated handover demonstrations},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Twu31/so101_hand_blue_napkin}
}

Recorded with LeRobot on SO-ARM101 hardware.