X-Humanoid, the Beijing Humanoid Robot Innovation Center, has released a general whole-body control model called Hi-WBC through its embodied-intelligence platform. The company announced the model on November 2 and provided further details in a Chinese press release on November 7.
Hi-WBC, which stands for Whole-Body Control for Humanoid Intelligence, is designed to treat human movement references as a guide rather than a script to be reproduced exactly. According to X-Humanoid, the model combines these references with robot dynamics and environmental constraints. This allows a robot to maintain the intent of a movement while adjusting its posture, joint motion, and support based on the specific load, external forces, and terrain it encounters.
The company identifies three specific capability areas for the model. To ensure robots can acquire new skills without losing previous ones, Hi-WBC utilizes a framework based on policy distillation and a mixture-of-experts. In this system, specialist expert policies learn specific manipulation tasks or movements, which are then distilled into a general policy.
For physical interactions involving heavy loads, the model uses a contact-aware learning approach. This method estimates external forces by analyzing motion states and joint currents, combining that data with reinforcement learning and short-horizon dynamics optimization. To navigate uneven ground, a perceptive control approach uses terrain geometry to determine viable footholds. This allows the robot to adjust its body posture, leg swing, and foot placement on obstacles, steps, and slopes. The release states the robot can decide how and when to clear an obstacle before resuming its walk.
X-Humanoid shared two demonstration videos to showcase the model. One video features the Tiangong robots in an all-round five-event competition at a robot sports event. X-Humanoid claims the Tiangong robots won gold and silver medals, noting they were the only bipedal humanoid to win gold in the scenario events. The second video shows several robots tidying a meeting room in one continuous shot, performing tasks such as moving chairs, shifting a whiteboard, lifting a trash bag, and squatting to pick up paper.
