RedirectMimic is a training method that lets a humanoid robot skip chosen movements from a reference motion while still following the rest. It trains on an edited reference, while the controller still sees the original.
In a Unitree G1 simulation, using a controller pre-trained on 7,509 motion clips, the overhead-arm movement still appeared 2.7% of the time, against 58.9% when rewards were zeroed and 32.8% with anti-tracking. The method reached 94.2% joint success, counting suppression, no falls and keeping the rest of the motion.
On a speed-based edit it had the lowest replacement error, and it tracked unedited motions about as well as the original controller.
Background
RedirectMimic
Humanoid motion editing method
A training method that replaces selected parts of a humanoid's reference motion, such as arm positions, while keeping the rest.
