Humanoid Chair Mobility Moves Beyond Standing

Single-source brief: An arXiv paper reports seated, omnidirectional motion on a passive chair with casters.

What Changed

Researchers report a learned control policy for a Unitree G1 humanoid on a mobile chair.

The chair is passive and uses casters. The robot must move itself and the chair.

The work targets seated loco-manipulation. That means moving while seated, then potentially working from that position.

According to the arXiv paper, the robot uses intermittent foot pushes against the floor. Its pelvis contact with the seat is not fixed.

How the System Works

The team added a passive-chair model to a standard standing control environment.

They also added seated-state rewards and contact settings for the task.

The policy receives joint-position feedback and velocity commands. It does not use contact sensing or chair-state inputs.

The paper says the policy was trained without motion-imitation rewards.

In simulated random-command tests, policies tracked omnidirectional commands through nearly all 20-second runs.

The strongest seated policies sometimes beat a standing policy on velocity tracking, according to the paper.

Efficiency and Control Limits

Foot-slip regularization reduced cost of transport, the paper reports. It also raised tracking error.

Some policies using only that setting settled into stationary behavior. Adding symmetry regularization or a command curriculum avoided that failure.

Motion direction also mattered. The reported energy cost ranked backward below lateral, with forward motion far higher.

Deployment Reality

The authors report zero-shot transfer from simulation to a Unitree G1. They say it produced omnidirectional seated movement.

This is a research result, not a confirmed deployment. No independent confirmation was supplied.

Key unknowns include real-world runtime, payload, chair compatibility, safety behavior, and performance outside the reported setup.