Motion Capture Data for Humanoid Motion

The buyer's problem

Whole-body control for a humanoid needs reference motion: walking, reaching, crouching, carrying. Teleoperating a humanoid for every posture is slow, and web video gives noisy pose estimates. Optical mocap offers cleaner kinematics from people, but the data belongs to human bodies with human proportions, so each use needs a mapping to the robot. Buyers ask where to find such data, how to license it and what to specify.

What published work did with human motion data

• AMASS unified 15 optical marker-based mocap datasets under the SMPL body model with a fitting method called MoSh++, spanning over 40 hours, more than 300 subjects and more than 11,000 motions (Mahmood et al., 2019). It is a research archive; commercial terms need a separate check.
• SMPL-X extends the body model with fully articulated hands and an expressive face, learned from thousands of 3D scans, and its fitting method estimates it from a single RGB image (Pavlakos et al., 2019). Models and code are available for research, so license terms are a diligence item.
• HumanPlus trained a low-level policy in simulation with reinforcement learning on 40 hours of human motion data, then shadowed a human's body and hands in real time from a single RGB camera on a customized 33-DoF, 180 cm humanoid; skill policies learned from up to 40 demonstrations per task reached 60 to 100 percent success (Fu et al., 2024).
• Humanoid-X mined internet video, captioned it, retargeted human motion to humanoids and trained a text-to-action model, producing more than 20 million humanoid poses with text descriptions (Mao et al., 2024).

These are results from specific systems on specific robots. They show that human motion data can feed humanoid pipelines; they do not set the amount, quality or format you need.

Sourcing routes

Mocap studios and specialist sensing providers supply the middle rows; this kind of data comes from studios, wearable rigs and research archives, not from ordinary business footage archives. Compare against other routes in existing assets, new collection or specialist providers. Pose estimated from workplace video is covered in full-body pose from workplace video; for distributed contact see tactile sensing data; hands are treated in 3D hand pose data.

Retargeting and the embodiment gap

Retargeting maps human joint motion onto a robot with different link lengths, joint limits, degrees of freedom and mass distribution. Feasible human poses can be infeasible or unstable for the robot, so a retargeting step usually adds constraints for joint ranges, self-collision, foot contact and balance, often followed by a simulation-trained tracking controller as in HumanPlus. Hands usually need their own mapping; see retargeting human hand demonstrations. The broader topic is in the human-to-humanoid embodiment gap.

When you buy mocap, ask for the raw marker data and the fitted skeleton, not just a retargeted result tuned to someone else's robot.

Formats and force channels

C3D is a long-standing biomechanics format: a binary file with a 512-byte header, a parameter section and a data section, storing 3D point and analog data such as force plates and EMG as 16-bit integers or 32-bit floats; the basic file stores 255 points per frame, extendable (C3D.ORG). Because analog channels can sit beside markers, a studio with force plates may deliver ground reaction forces synchronized with motion. See force/torque data for contact-rich tasks for what that can and cannot support, and sensor synchronization requirements for timing checks.

Illustrative mocap delivery package

# Illustrative - synthetic values
package_id: mocap_pkg_A
capture: {system: optical_markers, marker_set: "full_body_plus_hands", frame_rate_hz: 120}
subjects: [{id: performer_A12, height_cm: 175, release_ref: REL-0012}]
motions:
  - {clip_id: m_0001, task_label: "carry_box_two_hands", duration_s: 14.2}
formats: [c3d, bvh, fitted_body_model_params]
analog_channels: ["force_plate_1", "force_plate_2"]
props: {box_dims_cm: [30, 20, 15], mass_kg: 2.0}
calibration_report: mocap/calib_2026-04-02.pdf
labels: [task, phase, contact_events]

What mocap data cannot support

Mocap alone supplies no robot actions, no camera views at the robot's mounting, and no object or workplace appearance unless the studio includes it. Studios are controlled spaces, so there is limited realism for cluttered work sites unless capture takes place on site. Marker occlusion by objects, hands or clothing creates gaps. Data from performers does not reveal balance strategies a specific robot needs, and one performer's style may bias a dataset. Retargeted motion remains a reference, validated by simulation and hardware tests. Training a humanoid's whole-body controller still needs a tracking pipeline, and published pipelines such as HumanPlus add simulation training and robot-collected data.

Permissions and performer releases

Motion capture records identifiable body movement, and may include face scans or voice. This is not legal advice; check these items with counsel.

• Does each performer release cover commercial AI training, derived robot motion and sublicensing to a buyer?
• Can a performer withdraw consent, and what happens to the clips?
• Do releases cover hand scans or facial data if captured?
• Are any props or branded items third-party content?
• For archives, are the license terms for commercial machine-learning use clear, and who is the licensor?

See worker consent and identifiable people for related diligence.

Questions to ask suppliers

1. Which capture system, marker set and frame rate were used, and can you deliver raw markers?
2. How many distinct performers appear, and how were they selected?
3. Do performer releases explicitly cover AI and robot training and derivative motion?
4. Which skeleton or body model do the delivered files use, and can you share fitting residuals?
5. Are force plates, EMG or other analog channels recorded and synchronized?
6. Can you capture tasks with our objects and layout, on site or in studio?
7. How do you label contact events and task phases?
8. What share of frames contains marker occlusion or reconstruction gaps?
9. Do you deliver C3D or equivalent open formats?
10. What retargeting work, if any, is included, and for which target skeleton?

Discuss your requirements

If you need whole-body motion data for a specific humanoid, describe the motions, skeleton and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates license and delivery. If no existing asset fits, SourceX can scope a commissioned collection of new recordings. A request does not guarantee a match. Start at the humanoid training data guide or the procurement hub.

Discuss your humanoid training-data requirements with SourceX

Sources

• AMASS: Archive of Motion Capture as Surface Shapes — AMASS unifies optical marker-based mocap archives under a body model.
• Expressive Body Capture: 3D Hands, Face, and Body from a Single Image — SMPL-X body model with articulated hands.
• Learning from Massive Human Videos for Universal Humanoid Pose Control — Humanoid-X pipeline from human video to retargeted humanoid poses.
• HumanPlus: Humanoid Shadowing and Imitation from Humans — HumanPlus trains on human motion data and shadows humans on a humanoid.
• C3D file format (C3D.ORG - The biomechanics standard file format) — C3D stores 3D point and analog data such as force plate channels.