Beijing Humanoid Unveils Hi-WBC Whole-Body Control Model

Leaderobot.com reported on Nov. 9, 2026, that Beijing Humanoid officially unveiled its Hi-WBC whole-body control model on the “Huisi Kaiwu” platform on Nov. 2. The report said demonstrations showing the model’s control capabilities and applications were also released.

According to Leaderobot.com, Hi-WBC explores a method of estimating external forces from joint-state information and incorporating them into whole-body control. Because the model is intended for tasks involving loads or contact, evaluation conditions and safety on actual hardware will be important to determining whether its announced functions translate into real-world capabilities.

The focus of the report is not simply the introduction of a model that reproduces movements, but its positioning of Hi-WBC as a model that connects whole-body control with real-world tasks. However, this is an account of the announcement by Leaderobot.com; it does not show that the model’s performance or practicality has been independently verified. The excerpt describes demonstrations but provides no comparable figures for success rates, load conditions, terrain range or number of trials. The announced control concept should therefore be distinguished from the ability to perform work reliably in the field. The report describes a combination of MoE and policy distillation as a way of incorporating policies for specific skills into a general-purpose policy. This suggests that how well existing capabilities are preserved when skills are added will be one measure of the model’s value. The report also describes a method that estimates external forces from joint currents and joint-motion states and reflects them in control. This is not a matter of the control model alone: consistency with the robot’s sensor data, joint responses and dynamics model will also be an implementation requirement. The excerpt does not specify the hardware or whether the model can be transferred to different robots. From an industry perspective, the next focus will be how motion policies are evaluated on actual hardware and how safety and reproducibility are verified after new skills are added, rather than the model’s announcement itself. For tasks involving loads or external forces in particular, evaluation must consider not only successful movements but also behavior in the event of failure and stopping conditions. The excerpt does not make clear under what conditions the reported demonstrations were conducted or how well the model handles sustained real-world work. Items to watch include the target hardware, quantitative evaluations by task, the scope of data used for training and evaluation, safety measures against disturbances, and whether the model is being used in real operations. Until these points are clarified, “application to real-world tasks” is best treated as a goal stated at the time of the announcement. See the original report for details.