Humanoid robots must coordinate locomotion and manipulation, but updating one often forces retraining of the other. A new arXiv preprint argues that changing a robot's manipulation policy should not require retraining the controller that maintains balance. The proposed framework, DUET, decouples upper- and lower-body execution so the two can be developed independently.
DUET treats manipulation as a separate module from the balance and locomotion controller. This means a robot could learn a new grasping or tool-use skill without destabilizing the lower-body control that keeps it upright. The authors describe the approach as task-agnostic, suggesting it could apply across different manipulation tasks without reworking the whole control stack.
The abstract does not include experimental results or comparisons to other methods, so the claims rest on the proposed architecture rather than demonstrated performance. Still, the idea of modular control could simplify future humanoid development, making it easier to swap skills in and out as needed.