Teleoperation and live control
Wear the suit, stream human motion straight to the robot, and record every demonstration as training data in the same session.
Control now, training data for later
Three properties make the difference between a demo and a control loop you can actually train on.
High-fidelity data
A scientifically validated biomechanical model with full-body kinematics at production accuracy. The robot follows a clean, physically correct signal rather than a noisy proxy, so the same stream is worth keeping as training data.
Low latency control
End-to-end streaming fast enough to close the loop, including contact-rich and dexterous work. Latency stays predictable rather than drifting as the session runs.
Always-on streaming
Wireless WiFi, a wired tethered option and on-body buffering mean the stream survives interference. If the link drops, the buffer covers the gap instead of ending the take.
End-to-end latency
Capture rate
Wireless range
On-body buffer
How a teleop session runs
Four stages, start to finish. The same session produces both the control stream and the training data.
Suit up and calibrate
The operator puts on the suit and runs a short calibration. Setup is minutes, not hours, with one-click sensor mounting and no cameras or markers to place.
Stream to the robot
Motion streams to the humanoid over WiFi or a wired tether, retargeted onto your robot. The operator sees the result and adjusts in the loop.
Record while you control
Every demonstration is recorded on-body at full rate at the same time. Controlling the robot and building the dataset are one action, not two.
Retarget and train
Take the recorded demonstrations into your training pipeline through native exports, or reprocess them at higher quality afterwards.
Tasks the operator can drive directly
Full-body control, not just arms. If a human can demonstrate it wearing the suit, the humanoid can be driven through it.
Locomotion and stance
Walking, turning, weight shifts and crouched movement, with ground-truth foot contact and centre of pressure feeding the balance controller instead of an inferred estimate.
Manipulation and grasping
Reaching, grasping and tool use. Add Manus or StretchSense gloves and per-finger degrees of freedom stream in the same loop, which is what dexterous work requires.
Contact-rich tasks
Assembly, insertion, pushing and pulling, where the useful information is in how force and posture change on contact rather than in the trajectory alone.
Whole-body load handling
Lifting and carrying with the load in the loop. Segment scaling and anatomical modelling keep the kinematics valid across operators and body types.
Hardware, software and integrations
The three layers that make the teleoperation stack work end to end.
Motion capture suit
Full-body humanoid-grade capture with integrated WiFi streaming, a wired tethered option, on-body recording and hot-swappable batteries. Modular body set-ups and one-click sensor mounting keep changeovers fast.
Xsens Humanoid Software / SDK
Off-the-shelf for Linux and Windows, or in SDK form. Runs natively on ARM64, directly on the humanoid onboard compute, so there is no Windows bridge in the control loop.
Streams and retargeting
Native ROS 2 publisher today, with a native NVIDIA Isaac plugin and further simulator and middleware integrations rolling out. Retargeting runs through XsensTargeter or a research partner.
Control the robot and build the dataset at once
Most teleop setups make you choose: drive the robot, or record clean data. Here the control stream and the recording are the same signal.
The live stream
Retargeted motion goes to the humanoid in real time, so the operator closes the loop by eye and by feel. This is the part that gets the task done today.
The recorded take
The same motion is written on-body at full capture rate, untouched by the wireless link. This is the part that trains the policy tomorrow.
One clock, one model
Both come from the same biomechanical model at the same timestamps. No re-syncing, no second capture session, and no mismatch between what the robot did and what the dataset says it did.
Expand your teleoperation setup
Layer best-in-class third-party hardware on the same capture loop. Same session, same clock.
Finger tracking
Manus or StretchSense gloves add full per-finger degrees of freedom. For grasping, tool use and contact-rich manipulation this is not optional, it is the task.
HMD integration
Pair the suit with a VR headset for first-person control and immersive review. Meta Quest and Pico VR support coming soon.
Object tracking
Track tools, payloads and workspace objects alongside the operator with HTC Ultimate Tracker or Vive, so the robot has the scene and not just the body.
Video sync
Time-locked multi-camera video alongside the motion stream. Vision and biomechanics on one clock, ready for VLA training and annotation.
Absolute positioning
Lock the suit to a global reference frame so the session is spatially grounded for interaction and scene-aware tasks.
Not sure what you need?
Tell us the task you want the humanoid to perform and we will map it onto the right setup and modules.
Does it work with your robot?
Retargeting is what turns human motion into robot motion. We maintain our own retargeter and work with the research groups building the rest.
Supported humanoids
Unitree, AgiBot, Figure, Apptronik, Boston Dynamics, Agility, 1X, NEURA, Kepler, Fourier, UBTech, XPeng and more, with the current state per platform kept in the compatibility matrix.
XsensTargeter and partners
Our own XsensTargeter, plus support for research retargeters including GMR (Stanford), Westlake and MPI (Max Planck). Use ours or bring your own.
Not listed yet?
New platforms are added continuously. Tell us which humanoid you run and we will confirm where it stands.
Ready to put an operator in the loop?
Tell us which humanoid you run and what you want it to do. We will configure the suit, the modules and the retargeting path for your program.