HugWBC

A whole-body controller where one policy lets a humanoid walk, run, jump, and hop, with adjustable step frequency and foot-lift height.

HugWBC was released in February 2025 by Shanghai Jiao Tong University (Weinan Zhang's group) together with Shanghai AI Lab (Jiangmiao Pang) and others, published at RSS 2025. Most humanoid walking controllers only know one fixed gait, with no adjustable parameters and little room to extend. HugWBC designs a unified command space: besides velocity, it can also specify gait (walk/run, jumping, standing, hopping on one foot), step frequency, foot-lift height, body height, waist rotation, and body pitch. Training adds a symmetry loss, and uses “intervention training” to let the upper body be taken over by an external signal, such as teleoperation, which supports walking and manipulating at the same time. The policy is trained in Isaac Gym and deployed on a Unitree H1, with code open-sourced.

ExampleA remote control makes the H1 walk at a specified step frequency and foot-lift height, while the operator simultaneously takes over its arms through teleoperation to carry a box.

Also called

A Unified and General Humanoid Whole-Body Controller for Versatile Locomotion

Related

Whole-Body Control · Learning-Based Whole-Body Control · Gait · Loco-manipulation · Unitree H1 · HOVER

Sources

HugWBC (arXiv 2502.03206)
HugWBC project page
apexrl/HugWBC (GitHub)

As of

2025-04