Humanoid Training · Part of The Humanoid Group

Teleoperation for Data

Before a humanoid can learn a skill from people, someone has to drive it through that skill many times. This guide compares the rigs used for teleoperated data collection and sets out what separates a useful demonstration from a wasted one.

By Arjun Rao · Updated October 2026

Three families of teleoperation rig

Most teleoperation set-ups fall into three groups. VR headsets track the operator's head and hands and send the robot's camera views back to their eyes; Open-TeleVision, from UC San Diego and MIT, streams stereo video from cameras that move with the operator's head, which makes depth easier to judge during fine work. Leader-follower rigs give the operator a small replica arm or a pair of instrumented handles whose joints map onto the robot's own. Exoskeleton-style rigs strap to the operator's arms or whole body and capture joint angles directly, and some feed force back to the person. A survey of humanoid teleoperation by Darvish and colleagues places all of these within one architecture of operator, interface, communication link and robot.

Matching the rig to the task

The right rig depends on what you need to record. Fine two-handed work, such as seating a connector or folding cloth, rewards interfaces that capture finger motion and give the operator a clear view of the contact point. Whole-body tasks, such as squatting to lift a crate, need a rig that maps the operator's posture onto the robot's legs and torso without fighting its balance controller. Delay matters for every option, because a lagging video feed makes operators slow and hesitant, and that hesitation ends up in the recordings. Before buying hardware, run a short trial on your hardest task and watch where operators struggle. Ask the robot maker which rigs it already supports, since custom mappings take engineering time.

What a good demonstration looks like

A good demonstration finishes the task the same way whenever the situation is the same, and adapts sensibly when the situation changes. A Stanford-led study by Mandlekar and colleagues, published in the CoRL 2021 proceedings, found that the results of offline learning methods depended on the quality of the human demonstrations they were given, so sloppy runs are not harmless padding. Mark the start and end of every episode clearly. Spread the recordings across the real range of object positions, lighting and clutter. Include recoveries from small slips, because a robot that has only seen perfect runs has no example of getting back on track. Discard runs where the operator paused to think or the video feed dropped out.

Running a collection session well

Treat data collection as a production process with checks, not a free-form recording day. Write a short script for each task that fixes the starting layout, the goal and the moves that count as failure. Rotate several trained operators so the dataset is not shaped by one person's habits, and give newcomers practice time before anything is saved. Plan breaks, since headsets and body-worn rigs tire people quickly and fatigue shows up as jerky motion. Log the robot, software version, operator and scene with every episode so you can filter later. Review a sample of episodes each day; catching a calibration drift on the first morning costs far less than discovering it after a week of recording.

Sources and further reading

Teleoperation of Humanoid Robots: A Survey — Istituto Italiano di Tecnologia, Inria, IHMC, AIST and partners (arXiv).
Survey of humanoid teleoperation: the parts of a system, the interfaces operators use and where the method is applied.
What Matters in Learning from Offline Human Demonstrations for Robot Manipulation — Proceedings of Machine Learning Research (CoRL 2021), Stanford University and partners.
Peer-reviewed study of six offline learning methods, showing how results depend on the quality of human demonstrations.
Open-TeleVision: Teleoperation with Immersive Active Visual Feedback — UC San Diego and MIT (arXiv).
Open VR teleoperation system giving operators stereo vision that follows their head while their arms and hands drive a humanoid.

Common questions

What should I look for in a VR headset for robot teleoperation?

Look for reliable hand tracking, a comfortable fit for sessions that last hours, good pass-through or stereo video, and software your robot maker already supports. Test it with your own robot and network before buying in bulk, because integration problems often matter more than the headset's own specification.

Can teleoperated data be reused on a different humanoid?

Sometimes. Reuse is easiest between robots whose arms, hands and camera placement are similar. Larger differences mean the recorded motions must be retargeted, and some skills may not carry over at all, so check with both makers before planning to share datasets between platforms.

Learning From Demonstration — How humanoids learn skills from human demonstrations and teleoperation data.
Simulation & Sim-to-Real — Why humanoids train in simulation first, and what it takes to cross the gap to the real world.
Evaluation & Readiness — How to judge whether a humanoid is safe, reliable and ready for real work.
Preparing Your Team — The roles, skills and habits your organisation needs before humanoids arrive.
Humanoid Training Services — When to outsource robot data collection, teleoperation and evaluation, and what to ask a provider.
How to Train a Humanoid — A plain-language walkthrough of teaching a humanoid a new task, from demonstration to testing.
Teleoperation for Data — VR, leader-arm and exoskeleton rigs for recording humanoid demonstrations, and what makes a run worth keeping.
Learning to Walk — How reinforcement learning teaches humanoids to walk and balance in simulation before they step out for real.
Training Data and Privacy — Recording people at work or at home to train robots: consent, data protection law and careful data handling.
Choosing First Tasks — How to pick a humanoid's first skills, set pass marks and budget for the demonstrations each one needs.
Safety While Training — Zones, stops and supervision that keep people safe while a humanoid is being taught, tested and debugged.
Latest updates — Teaching robots now takes data operations, simulation and careful testing. These figures and dated updates show where training data comes from, how fast robots learn new tasks and how results are checked.
About us — What we cover, how we source every figure, and who publishes this site.

Visit The Humanoid Group

People also search for teleoperation data collection, humanoid teleoperation, VR teleoperation robot, exoskeleton teleoperation, leader-follower arms and robot demonstration data.