BRACE teaches the Unitree G1 humanoid to hold force and stand on a slope without a height map

A team of researchers has presented BRACE, a whole-body motion tracker for the Unitree G1 humanoid robot. It can apply a specified force with its hands, compensate for an external load, and stand on a slope. The operator only needs to send an ordinary trajectory recorded on a flat floor with no load; the robot accounts for the surface incline and the weight of the object by itself. The control loop contains no height map and no force sensor, only proprioception. The work is published in Hugging Face Daily Papers. The experiments were run in simulation and on a real robot.

The problem: the reference is recorded in an ideal world

Whole-body tracking has become the main way to control humanoids. The operator moves, and the robot repeats the motion. But reference motions are recorded on a flat surface with empty hands. The tracker that consumes them knows nothing about the forces of interaction with objects or about the incline of the support surface.

Existing controllers address only one side of the task. Force policies set the effort at the end effector but do not prescribe the pose of the whole body. Terrain-adaptive trackers treat a load as a disturbance to be suppressed, not as a wrench that can be controlled. As a result, a robot can either push on an object or avoid falling on a slope, but it cannot do both.

How BRACE works

The authors do not make the tracker absorb the load and the slope on its own. Instead, they rewrite the reference it follows. There are two mechanisms for this.

The first is terrain conformance. It lifts the foot contact points and the body root onto the local surface, so a flat trajectory becomes a trajectory for a slope.

The second is wrench transformation. It determines how far the hand must be shifted to produce the specified force. The calculation is done jointly with the shifts of the center of mass and center of pressure that this force causes. The result is bounded by the arm effort limits, and each teacher has its own limits.

Distillation comes next. The separate teachers, one for applying force and one for compensating it, are merged into a single student on flow matching using DAgger. The student receives only proprioception as input and handles the slope and the load on its own.

What is known about the results

The experiments were run on Unitree G1, in simulation and on a real robot. The abstract claims the ability to apply force, compensate for it, and cope with different terrain. It gives no quantitative metrics: no success rates, no force range, no limiting slope angles. The robustness margin of the approach can only be judged from the full text of the paper.

Key parameters from the description:

• platform: Unitree G1;
• training method: two teachers (exertion and compensation), distillation via DAgger;
• student architecture: flow matching;
• student input: proprioception only;
• not required: a height map and a measured wrench;
• command signal: a trajectory for a flat surface, including from a remote operator;
• validation: simulation and a real robot.

Practical significance

The main benefit of the approach is a simpler sensor stack. If a policy works without a height map and force-torque sensors, the platform is cheaper and has fewer sources of failure. It is also convenient for teleoperation: the operator does not have to compensate for terrain and payload weight by hand, and can work with ordinary motion recordings.

Several things are worth checking in an evaluation. What range of slopes and forces can the robot actually withstand? How does the system behave with latency in the remote-control channel? Does the approach transfer to humanoids other than the G1? Does it cope with uneven terrain beyond slopes, such as steps or soft ground? So far, the public data has no answers to these questions. Still, the scheme itself, which moves the complexity from the tracker into the reference, looks like a workable way to combine force control with stability on uneven footing.