MimicX: Policy-in-the-Loop Supervision Refinement for Video-Driven Humanoid Motion Tracking

MimicX. Policy-in-the-Loop Supervision Refinement for Video-Driven Humanoid Motion Tracking. By Shuaijun Liu and nine others, from the Hong Kong University of Science and Technology, Guangzhou. Published on the sixth of October. Its code repository is worked out: about two hundred source files, with a readme and a licence.

Human videos are a rich source of motion targets for humanoid robots. A person does something on camera, and the robot learns to track it. But a reference that looks plausible on screen can still fail, and keep failing, once it is executed under physics. Those failures are not just noise. They carry information, because they point to where the training supervision ought to change.

MimicX is a policy-in-the-loop framework built on that observation. It uses feedback from execution to refine video-driven humanoid motion tracking, so the policy's own failures become the signal for what to fix.

It begins from motion that has been reconstructed from video and retargeted to the robot. MimicX then localizes two things: the difficult transitions, and the body regions they affect. For those, it jointly adapts the tracking objectives and the reset curriculum used for policy continuation. Every candidate change is checked by repeated rollouts, and the framework selects improvements that put execution first, subject to guards that keep tracking from slipping.

Across four core video tasks, MimicX consistently improves both tracking accuracy and the Robust Execution Horizon, compared with the Fixed Reference baseline, which keeps the original supervision unchanged. Averaged over tasks, body-tracking error falls by twenty-five point seven percent. The execution horizon rises by two hundred and fifty-five point six percent.

Further studies on additional videos, motion references and collision scenes test the method beyond the core tasks. A variant called MimicX-HLoop speeds up the feedback through heterogeneous execution.

Here, a failed rollout is treated as supervision rather than as a dead end. It tells the system which part of the reference to refine, and repeated checking decides which refinement to keep.