ReActor: Physics-Based Motion Retargeting with Reinforcement Learning

Disney Research’s ReActor is a reinforcement-learning method for retargeting human reference motion to substantially different robot morphologies, including humanoids and quadrupeds. By integrating ph

The video presents the results of ReActor, which transfers human motion references to fit the morphology of robots. The researchers apply motion to humanoid robots and quadrupedal robots whose body proportions and joint structures differ from those of humans, showing that motion retargeting is not simply a matter of copying poses but of creating physically executable movement.

With existing methods, feet may slide across or appear to float above the floor, and body parts may penetrate one another. ReActor jointly handles the stage of adapting reference motion to the robot’s morphology and the training of a reinforcement-learning policy that tracks the motion through a bilevel optimization process. The required input is a specified semantic correspondence between selected rigid parts of the human and robot, reducing the burden of manually tuning parameters.

The paper presents applications on two types of humanoid robots and the ANYmal D quadrupedal robot. For the real Lima robot, a tracking policy trained on data converted by ReActor was deployed for validation. In an interactive environment, the study also presents a use case in which motion is retargeted to the robot’s form in real time while an artist edits the motion. The research was introduced as an ACM SIGGRAPH 2026 technical paper.

Maru