dadm022/lerobot_rgb_vla

LeRobot RGB VLA Dataset

A large real-robot manipulation dataset in LeRobot v3.0 format: 2,318 episodes, 2.49 million frames, 23 hours of teleoperated SO-follower arm demonstrations recorded with two synchronized RGB cameras.

Left to right: global (front) camera view, wrist camera view, and the front depth visualization. This dataset ships the two RGB streams; the depth panel comes from the same RGB-D recording session and is shown here to illustrate what the capture rig sees.

At a glance

Episodes	2,318
Frames	2,487,292
Duration	~23.0 hours @ 30 fps
Episode length	848–1,495 frames (~28–50 s), mean ~1,073
Tasks	62 recording sessions
Robot	so_follower (SO-100 / SO-101 family), 6 DoF
Cameras	front 1280×720, wrist 640×480 — both AV1, 30 fps
Format	LeRobot codebase_version v3.0
Size	51.1 GB
Modalities	RGB video + joint states + actions (no depth)

Preview

Nine sampled episodes from the front camera — the global view of the workspace used as the primary visual observation:

The same episodes from the wrist camera, which follows the gripper and captures the fine-grained contact phase of each manipulation:

Features

Key	Type	Shape	Notes
observation.images.front	video	720 × 1280 × 3	global workspace view, AV1
observation.images.wrist	video	480 × 640 × 3	gripper-mounted view, AV1
observation.state	float32	6	joint positions
action	float32	6	commanded joint positions
timestamp, frame_index, episode_index, index, task_index	—	—	LeRobot bookkeeping columns

Both observation.state and action use the joint order shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, gripper.pos.

Task strings are recording-session identifiers (for example i4_2-40-90) rather than natural-language instructions. Treat task_index as a session/condition label; if you need language conditioning, annotate on top of these groups.

Structure

data/chunk-{chunk:03d}/file-{file:03d}.parquet     # one file per episode, numeric columns only
videos/observation.images.front/chunk-000/file-*.mp4
videos/observation.images.wrist/chunk-000/file-*.mp4
meta/info.json                                     # schema, fps, path templates
meta/episodes/chunk-000/file-000.parquet           # per-episode lengths, offsets, video timestamps
meta/tasks.parquet
meta/stats.json                                    # normalization statistics

Video files are shared across episodes: each episode's segment is located through videos/<key>/from_timestamp and to_timestamp in the episode metadata, never by file boundaries.

Usage

Requires lerobot >= 0.4 (tested against the v3.0 dataset format used by 0.6.x).

from lerobot.datasets import LeRobotDataset

dataset = LeRobotDataset("dadm022/lerobot_rgb_vla")
sample = dataset[0]
print(sample["observation.images.front"].shape)  # [C, H, W]
print(sample["observation.state"], sample["action"])

Stream it without downloading the full 51 GB:

from lerobot.datasets import StreamingLeRobotDataset

dataset = StreamingLeRobotDataset("dadm022/lerobot_rgb_vla")

Provenance

Recorded, merged, and audited with Loop Engineering for VLAnything — an open-source LeRobot toolkit for capturing, merging, auditing, enriching, and iteratively improving VLA datasets. Project page: https://dataset-toolkit-vlanything.derek-xia.chatgpt.site/

This release is the RGB-only export of an RGB-D collection. Episode, frame, task, and video indices were rebuilt during merging, so all indices are dense and contiguous, and every episode's video segment length matches its action sequence exactly.

License

Released under the MIT License.