PhaseTrack: Physics-Based Motion Tracking via Phase-Guided Motion Generation

In this work, we introduce PhaseTrack, a novel phase-guided physics-based motion tracking method that enhances motion generation quality by leveraging deep phase features learned from a periodic autoencoder. PhaseTrack employs a novel integration of phase-conditioned sparse mixture of experts and hierarchical encodersparse mixture of experts architecture conditioned on phase features within a world model framework, enabling the effective extraction of information from highly abstract phase representations without compromising real-time inference performance. Additionally, we propose a detail encoder that captures high-frequency motion details from reference target states, ensuring the realism of the generated motions. Our comprehensive evaluations demonstrate that PhaseTrack outperforms state-of-the-art model-based motion tracking methods in terms of precision and motion fidelity.