BRACE: Adapting Whole-Body References for Force and Terrain Aware Humanoid Motion Tracking

BRACE. Adapting Whole-Body References for Force and Terrain Aware Humanoid Motion Tracking. By Sudarshan Harithas and seven others, from Multiply Labs, Brown University, and one other institution. Published on the fifth of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Operators now drive humanoid robots largely through whole-body tracking: give the robot a reference motion and it follows. The catch lies in where those references come from. They are recorded on level ground, by someone carrying nothing. A tracker following them therefore knows nothing about the forces the robot will have to exchange with objects, and nothing about the terrain under its feet. Existing controllers cover one side of this gap at a time. Force-capable policies can command a force at the end effector, but they prescribe no whole-body pose. Terrain-adaptive trackers handle uneven ground, but treat any load as a disturbance to reject, not as a wrench to command.

BRACE is a whole-body tracker meant to do both at once. It exerts commanded hand forces, and compensates for them, from diverse poses, while following a flat-ground reference across sloped terrain. Its central move is about where the work happens. Instead of leaving the tracker to absorb the load and the slope as errors, BRACE folds both into the reference itself, so the motion being tracked already accounts for them.

Two adjustments reshape that reference. Terrain conformance lifts the footholds and the root onto the local surface. A wrench transformation then resolves the hand displacement that produces a desired force, jointly with the shifts that force induces in the center of mass and the center of pressure. That resolution is bounded by arm-effort limits specific to each teacher. Pushing on something, in other words, is not treated as a matter for the hand alone. Where the hand goes, where the body's weight sits, and where the feet press against the ground are worked out together, before the tracker ever sees the motion.

There are two teachers: one for exertion, one for compensation. DAgger distills both into a single flow-matching student. The student runs on proprioception alone, with no height map and no measured wrench. In practice that means a teleoperator, even one in a remote location, can supply an ordinary flat-ground trajectory, and the robot works out the slope and the load on board.

The authors report extensive experiments on the Unitree-G1, in simulation and on the real robot. These are presented as demonstrating that BRACE can exert force, compensate force, and handle diverse terrain. The abstract reports these capabilities qualitatively and gives no figures for them.

Underneath the robotics is a choice about interfaces. The flat, unloaded reference stays what the operator provides, and is not asked to become anything richer. What changes is the layer between that reference and the body: a step that rewrites the motion for the ground and the force actually present, before a single learned student tracks it.