Scalable and General Whole-Body Control for Cross-Humanoid Locomotion

Learning-based whole-body controllers have become a key driver for humanoid robots, yet most existing approaches require robot-specific training. In this paper, we study the problem of cross-embodiment humanoid control and show that a single policy can robustly generalize across a wide range of humanoid robot designs with one-time training. We introduce XHugWBC, a novel cross-embodiment training framework that enables generalist humanoid control through: (1) physics-consistent morphological randomization, (2) semantically aligned observation and action spaces across diverse humanoid robots, and (3) effective policy architectures modeling morphological and dynamical properties. XHugWBC is not tied to any specific robot. Instead, it internalizes a broad distribution of morphological and dynamical characteristics during training. By learning motion priors from diverse randomized embodiments, the policy acquires a strong structural bias that supports zero-shot transfer to previously unseen robots. Experiments on twelve simulated humanoids and seven real-world robots demonstrate the strong generalization and robustness of the resulting universal controller.

Humanoid robots are being developed by many companies, but each robot has a different mechanical structure, morphology, and physical properties. As a result, most controllers today must be trained separately for each robot, which requires substantial time and effort and makes it difficult to scale to new platforms. In this work, we show that a single “brain” can be trained once and then deployed across many different humanoid robots without additional training. Our method, called XHugWBC, exposes the controller to a wide range of simulated robot designs during training and represents all robots in a common format. This allows one neural network to understand and control robots with very different shapes and dynamics. We validated XHugWBC on tens of thousands of simulated robot variations and on seven real humanoid embodiments from different manufacturers. The controller transferred successfully to all real robots in a zero-shot setting, meaning it worked immediately without any robot-specific retraining. To the best of our knowledge, this is the first work to extend a unified “brain” beyond simple locomotion to whole-body humanoid control, enabling complex whole-body tasks. This work suggests that future humanoid robots could share a universal control system, much like different computers can run the same operating system. Such a general-purpose controller could greatly reduce the cost and engineering effort required to develop and deploy new humanoid robots, accelerating their adoption in homes, factories, and public spaces.