01 / DUAL-ARM LIFT
Coordinated manipulation
Two arms, two grippers and a lift axis.
Object handling, assembly and everyday tasks.
- Tasks
- 169
- Hours
- 6,749
- Preview videos
- 08
ROBOT & HUMAN MOTION DATA / OCT 2026HIYIO · EMBODIED INTELLIGENCE
Real-world
motion data.
Human demonstrations.
Synchronized signals. Real motion.
Partner data · Privacy-edited technical preview
01 / REAL CAPTURELIVE-ROBOT RECORDING
Coordinated manipulation · Block sortingWatch the demonstration ↗
HUMAN DEMONSTRATION → ROBOT LEARNING
Two platforms. A broad task inventory.
From coordinated tabletop manipulation to tasks across a workspace.
351platform task records
12,787hours in the shared inventory
784,315recorded episodes in the inventory
02 / WHEELED HUMANOID
Mobile manipulation
Moving between task locations.
Handling objects in everyday environments.
- Tasks
- 182
- Hours
- 6,038
- Preview videos
- 03
Figures summarize the supplied inventory. Same-name tasks on different platforms are counted separately. The downloadable preview is a subset.
Watch the task. Read the action.
Eleven real demonstrations. Select a task, play the video or jump to an annotated action.
Video unavailable. Select the task again or refresh the page.
DUAL-ARM / A01
Block sorting
Videos are edited for privacy and compressed for presentation. Mobile examples currently include video only, without action-level annotations.
Find the next task.
Search the supplied inventory by task, platform or environment.
Loading inventory…Sorted by episode count
| TASK | ENVIRONMENT | PLATFORM | EPISODES | HOURS |
|---|
Environment categories support browsing. Same-name entries are combined within each platform; unresolved source labels remain marked for review.
One episode. Aligned modalities.
A field-level view of the dual-arm example, from raw capture to a training sequence.
MCAPRaw multimodal capture
→Align & convertShared episode and frame indices
→LeRobotParquet + camera streams
Check the schema and control conventions for each platform and delivery.
Dictionary + license ↓Start with a real episode.
Block sorting · Original numerical values, with identifying metadata removed.
LEFT ARM / JOINT 1
Joint position · radLeader command · action[0]Measured state · state[0]
Original, unshifted channels. Displayed every four frames; no timing correction applied.
1,187frames in one episode
- Sequence
- 30 fps / 39.57 s
- Action
- 17 dimensions
- Measured state
- 71 dimensions
The download contains the numerical frame table. Training use requires the corresponding vision, calibration and alignment checks.
The plotted values come from the downloadable Parquet. Command and measured state are separate signals; their difference is not an independently calibrated error metric.
See alignment in the image.
Two complementary views for inspecting geometry and spatial structure.
These are inspection examples. A projection overlay does not establish batch-level calibration accuracy; false-color depth is a visualization, not the raw depth values.
Quality you can verify.
Supplied acceptance references, to be checked against each delivery batch.
Vision & frame timing
RGB ≥ 1280 × 720 · Target 30 fps
Frame intervals: P1 ≥ 23.33 ms and P99 ≤ 43.33 ms, checked per episode and within rolling windows.Multimodal synchronization
Camera matching: |mean time difference| < 5 ms
Vision-to-state alignment delay ≤ 5 ms. Timestamp resolution alone does not prove synchronization accuracy.Measured signals
Follower sampling ≥ 195 Hz · Teleoperation side ≥ 62 Hz
Retain measured joints, grippers, poses, force/torque and timestamps. Do not substitute values inferred through inverse kinematics.Task completion
Review the action, the object and the final state.
Check grasp and placement outcomes, continuous motion, joint limits, modality completeness and privacy.01 / CAPTURE Logs, frequency, timestamps→02 / CONVERT Fields, indices, loadability→03 / REVIEW Task, result, privacy
These thresholds describe the supplied criteria, not a claim that the full inventory has been independently validated.
See the action in 3D.
Recorded trajectories drive the model alongside three synchronized camera views.
XYZ
From recorded signals to visible motion.
Load the model, trajectory and camera streams when you are ready.
Drag to orbit, scroll to zoom, and scrub the timeline to inspect motion. The source model geometry and recorded numerical trajectories are retained.
Human motion. Frame by frame.
Recorded sequences, with original joint hierarchies and motion channels preserved.
03 BVH sequences60 / 120 Hz capture16,988 frames
Open full viewer ↗HUMAN MOTION / 3D REPLAY
Read the movement.
Inspect the skeleton.
Play, scrub and orbit the captured motion.
Download each sequence as BVH.
Neutral sample IDs replace identifying filenames. Skeleton replay uses the recorded channels; robot retargeting requires a separate model and mapping.
One flip. From video to robot.
Webster flip: source video, extracted SMPL-X motion and T800 retargeting.
SOURCE VIDEOEXTRACTED SMPL-XT800 RETARGETING
- HUMAN EXTRACTION
- 164frames / 30 Hz
- T800 RETARGETING
- 273frames / 50 Hz
- ROBOT CHANNELS
- 25joints + root pose
Online-source footage processed for extraction and retargeting. Original rights remain with their owners. For copyright concerns, contact us for removal.
FROM DEMONSTRATION TO DATA REVIEW
Explore.
Inspect.
Validate.
Use real tasks and real signals
to begin a concrete data conversation.
01 / CSV↓02 / PARQUET↓03 / JSON↓04 / VIDEO + JSON↗05 / BVH↗06 / NPZ + CSV↗
Task inventory
Platform, environment, episode counts and hours.
Structured episode
1,187 frames · 17 action values · 71 state values.
Field dictionary
State, action, image streams and frame indices.
Annotated demonstrations
Select a task to download its preview and annotations.
Human motion capture
Recorded sequences with skeleton replay and BVH downloads.
Video to robot motion
Webster flip: extracted SMPL-X and retargeted T800 data.
MORE MOTION SAMPLES
Publisher-hosted CSVs · Chingmu source terms ↗
Local downloads include LICENSE.txt and source notices in ZIP packages.
Prepared from shared partner materials · October 2026.
Preview media is edited for privacy. Dataset scope and acceptance remain specific to each delivery.