so101_wm

Teleoperated manipulation data from an SO-101 arm, recorded for training visuomotor policies and action-conditioned video prediction models. The task is stacking and unstacking large 3D-printed nuts, deliberately demonstrated with varied grasps and varied placement locations rather than a single repeated motion.

177 episodes, 355,884 frames at 30 Hz, two synchronized camera views, released in LeRobot v3.0 format.

At a glance

Episodes	177
Frames	355,884
Duration	3.3 hours (67 s per episode on average)
Frame rate	30 fps
Tasks	2 (stack, unstack)
Robot	SO-101 follower arm (so_follower), 6 DoF including gripper
Cameras	2 x RGB, 640x480, AV1
Format	LeRobot codebase_version: v3.0
Size	7.53 GB
Split	train (all 177 episodes)
License	MIT

The task

Large 3D-printed nuts are stacked into a pile and unstacked again. The objects are big enough that there is no single correct grasp: the same nut can be picked from the side, from the top rim, or at different angles around its circumference, and placements land at different positions in the workspace across episodes.

This was intentional. A dataset where every episode executes the same trajectory is easy to fit and teaches a policy very little. Here the action distribution is genuinely multimodal conditioned on the observation, which is the regime where naive behaviour cloning tends to average incompatible demonstrations into a trajectory that satisfies none of them, and where diffusion policies and flow-matching action heads are supposed to earn their keep. If you want a small dataset for comparing unimodal against multimodal action decoders, that is what this is for.

Collection setup

• Robot: SO-101 follower arm, teleoperated through a leader arm. Recorded joint positions are the follower's, and actions are the commanded targets from the leader.
• Cameras: two RGB streams at 640x480, recorded in step with the 30 Hz control loop.
  • observation.images.fpv — wrist-mounted, moves with the end effector.
  • observation.images.left — static third-person view of the workspace.
• Control rate: 30 Hz, matching the video frame rate, so one frame corresponds to exactly one action.
• Encoding: video is AV1 in yuv420p, no audio, no depth.

Data structure

Each frame carries:

Key	Type	Shape	Notes
action	float32	(6,)	commanded joint positions
observation.state	float32	(6,)	measured joint positions
observation.images.fpv	video	(480, 640, 3)	wrist camera
observation.images.left	video	(480, 640, 3)	third-person camera
timestamp	float32	(1,)	seconds from episode start
frame_index	int64	(1,)	index within episode
episode_index	int64	(1,)	episode id
index	int64	(1,)	global frame index
task_index	int64	(1,)	task id, see meta/tasks.parquet

Both action and observation.state use the same joint ordering:

shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos,
wrist_flex.pos,   wrist_roll.pos,    gripper.pos

Layout on disk follows the LeRobot v3.0 convention:

data/chunk-{chunk:03d}/file-{file:03d}.parquet
videos/{video_key}/chunk-{chunk:03d}/file-{file:03d}.mp4
meta/info.json, meta/stats.json, meta/tasks.parquet, meta/episodes/

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("shubham4413/so101_wm")

print(ds.num_episodes, ds.num_frames, ds.fps)
sample = ds[0]
print(sample["observation.images.fpv"].shape)  # (3, 480, 640)
print(sample["action"])                        # (6,)

On older LeRobot releases the import path is lerobot.common.datasets.lerobot_dataset.

To load action chunks and an observation history, which is what diffusion policies and most VLAs expect:

delta_timestamps = {
    "observation.images.fpv":  [-1 / 30, 0.0],
    "observation.images.left": [-1 / 30, 0.0],
    "observation.state":       [-1 / 30, 0.0],
    "action":                  [i / 30 for i in range(16)],  # 16-step chunk
}

ds = LeRobotDataset("shubham4413/so101_wm", delta_timestamps=delta_timestamps)

Training a diffusion policy with the LeRobot CLI looks roughly like:

lerobot-train \
  --dataset.repo_id=shubham4413/so101_wm \
  --policy.type=diffusion \
  --output_dir=outputs/so101_wm_diffusion

Exact flags and entrypoint names move between LeRobot versions, so check against the version you have installed.

Decoding note

The videos are AV1. Decoding needs a reasonably recent torchcodec or an FFmpeg build with an AV1 decoder (libdav1d). If you get empty frames or a decoder error, that is almost always the cause rather than anything wrong with the files. Transcoding to H.264 locally is a valid workaround if your stack is older.

What this is good for

• Training and comparing action decoders on multimodal demonstrations: diffusion policy, flow matching, action chunking transformers.
• Fine-tuning small VLAs such as SmolVLA or π0 on a real, non-simulated, non-toy manipulation task.
• Action-conditioned video prediction and latent world models. The _wm in the name is the reason the dataset exists: two viewpoints at a fixed rate with aligned actions is the minimum you need to train something like V-JEPA 2-AC and then compare planning in a learned latent dynamics model against a policy trained directly on the same episodes.

Limitations

Worth knowing before you build on this:

• Single environment, single operator. One workspace, one lighting condition, one person teleoperating. Expect a policy trained on this to generalize poorly to a rearranged scene, and treat it as a benchmark for fitting multimodal demonstrations rather than for testing robustness.
• No force or torque signals. State is joint positions only. Contact-rich behaviour has to be inferred from vision and proprioception.
• No depth. Both streams are RGB.
• No failure episodes. Demonstrations are successful attempts, so there is nothing here for training a recovery or a reward model.
• No held-out split. Everything is in train. Hold out episodes yourself, and split by episode_index rather than by frame, or frames from the same episode leak across the boundary.
• Success is not labelled per frame. There is no explicit outcome annotation beyond the task index.

Citation

@misc{nagar2026so101wm,
  title  = {so101_wm: Teleoperated SO-101 stacking and unstacking demonstrations},
  author = {Nagar, Shubham},
  year   = {2026},
  url    = {https://huggingface.co/datasets/shubham4413/so101_wm}
}

License

MIT. Use it for whatever you like. If you train something interesting on it, I would genuinely like to hear about it.