Retargeting large human motion capture datasets to different humanoid robots is usually slow and needs manual tuning. The proposed FRDV pipeline splits the problem into a fast kinematic stage and a separate dynamics validation stage. The kinematic stage uses a custom Sequential Quadratic Programming solver with Real-Time Iteration and parallel chunk-based processing, reaching up to 9,156 frames per second—roughly 136 times faster than GMR and ProtoMotions.
Robot-specific constraints are folded into the optimization, which the authors say reduces common retargeting artifacts such as ground penetration and self-collision. For motions that need extra verification, a contact-implicit iLQR controller tests the retargeted reference under full rigid body dynamics. This step flags dynamically infeasible motions and cleans up contact artifacts like foot floating and skating.
To measure quality, the team trained motion tracking reinforcement learning policies with several state-of-the-art pipelines and compared tracking fidelity and success rates. They also demonstrated sim-to-real transfer on a Unitree G1 humanoid. Since this is a single arXiv paper, there are no independent results to compare; the claims are the authors' own.