Combining motion priors and motion tracking for both accuracy and adaptability

Determine algorithmic frameworks that simultaneously achieve accurate imitation of reference motions and broad adaptability to diverse deployment conditions in humanoid control, effectively combining the strengths of motion-prior-based reinforcement learning with motion-tracking methods without requiring large-scale training motions or test-time target trajectories.

Background

The paper distinguishes two major paradigms for motion-based humanoid control: motion priors integrated into reinforcement learning, which tend to enable broad adaptation but can sacrifice imitation accuracy; and motion tracking methods, which achieve high-fidelity reproduction of reference motions but typically require large datasets and often a test-time target motion. Bridging these paradigms to obtain both accurate imitation and wide adaptability is identified as a central difficulty in the field.

OmniH2O is proposed as an initial step towards this goal, aiming to adapt from a single reference motion while preserving motion patterns. The statement clarifies that achieving the full combination of strengths remains unresolved, motivating further research into unified approaches that do not rely on extensive data or reference trajectories at deployment.