The problem sounds simple. A person moves, and a humanoid follows. The bodies have different proportions and joint limits, and the robot has to stay physically stable. A useful system needs to translate the movement into actions the machine can execute.
The paper
Learning Human-to-Humanoid Real-Time Whole-Body Teleoperation, submitted March 7, 2024 by Tairan He and colleagues, presents H2O. The authors describe estimating human motion from an RGB camera, selecting feasible retargeted motions, training a controller in simulation, and transferring it to hardware. They report demonstrations including walking, turning, and pushing. Read the paper and abstract ↗
Three ideas in plain language
- Retargeting
- Mapping movement from one body onto the joints and proportions of another.
- A control policy
- A rule, often learned, that turns observations and desired behavior into actions.
- Simulation-to-reality transfer
- Developing a behavior in a simulated environment, then putting it to work on physical hardware.
Why this matters to readers of robot news
Live human guidance and learned robot control can appear in the same system. The operator may supply the movement target while the software solves part of the problem of executing it. That division of labor deserves explanation whenever a performance is described.
A useful follow-up question is whether the robot later performs a task from a goal, continues following a human, or switches between modes. The presence of machine learning alone does not answer that question.
How to assess the result
Begin with the authors’ stated task and evaluation conditions. Then look for the range of motions, hardware used, failure cases, and comparisons in the full paper. For operational applications, ask what happens during extended use and when the operator’s motion falls outside the tested range.
Sources & reading
Sources reviewed November 5, 2026. Company claims and author-reported research are attributed in the text.