LYRIC: Language-Driven Physics-Based Character Control for Contact-Rich Whole-Body Object Interaction

arXiv cs.GR details LYRIC, a generative flow-matching controller for language-driven, physics-based character control in contact-heavy object interactions. The goal is to let simulated humanoids follow a natural-language instruction plus a sparse terminal object goal, instead of requiring a full kinematic reference for every pose and contact.

The pipeline is split into a task-level planner and an action generator. The planner predicts short-horizon object and root trajectories, while the action generator resolves whole-body motion and contacts in closed loop. To make training work with imperfect mocap, the authors use geometry-conditioned interaction rewards and relax reference tracking near hand-object contact.

For developers, the interesting part is the combination of semantic intent and physically plausible interaction. In OMOMO evaluation, the tracker reached 64.3% success versus 53.2% for an InterMimic reimplementation, and the unified policy hit 76.5% on the full dataset. On the held-out split, LYRIC reached 90.3% task success versus 74.2% for the strongest matched kinematic-planner baseline, with better semantic alignment and motion quality.

The controller also supports test-time object-waypoint guidance without retraining, and the qualitative results show robust contact handling plus zero-shot transfer to novel object shapes. That makes it relevant for anyone exploring AI-assisted animation, embodied agents, or game characters that need to manipulate props in a way that still looks grounded in physics.