Growing a humanoid controller continuously: unified multi-skill control on top of a stable locomotion substrate
Year: 2026
Type: article
Abstract: Humanoid controllers deployed in the field are asked to gain new skills, such as get-up recovery or jumping, without retraining from scratch and without degrading their validated locomotion. We treat a deployed locomotion controller as a reusable control substrate on which new skills are grown rather than rebuilt, and present GRAFT (Growing skills on a Reusable, Anchored Foundation via Transfer), a stage-wise expansion framework whose training components play distinct roles. Warm-starting each stage from the deployed controller accelerates new-skill acquisition. Structured replay keeps base locomotion tasks sampled and optimized at every stage and maintains base-task performance throughout expansion. A locomotion-anchoring loss keeps the additive correction that new skills impose on the locomotion actions small: it reduces the locomotion action deviation dose-dependently and holds that deviation at its pre-conflict baseline even under a backward-walking skill trained to conflict directly with the forward gait. On a simulated Unitree G1, a single skill-conditioned policy carries the robot through walking, push recovery by getting up, and a full takeoff-air-landing jump chain, and extends to stylized dance and kungfu, from one policy file with no state machine.
Source: Advanced Robotics
Institution: Waseda University
Language: English
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7220994591