Physics-Based Motion Tracking of Contact-Rich Interacting Characters

How to read this

Category

physics-based motion tracking, reinforcement learning controller for multi-character interaction

Contributions

• Proposes a progressive mixture-of-experts controller, built on the progressive neural network framework, that tracks two-character contact-rich interactions such as boxing, pushing, and grappling in a single policy
• Automatically routes training samples to experts by tracking error and difficulty rather than requiring hand-scheduled task partitioning, unlike the original progressive neural network formulation
• Eliminates the need for the multi-stage, skill-specific training pipelines used by prior two-character interaction work
• Demonstrates robust tracking on the InterHuman dataset with stability under perturbations from obstacles or altered next-frame poses, and analyzes what different experts specialize in

Context

The work extends single-character physics-based motion tracking, citing Luo et al.'s line of AMASS-scale tracking controllers, into two-character contact-rich settings, and explicitly builds on and critiques prior two-character work: Won et al.'s task-specific boxing and fencing controller and Zhu et al.'s multi-stage boxing synthesis, arguing both are too narrow or too costly to train. It borrows the progressive neural network idea from reinforcement learning to solve the graphics-specific problem of force transfer through contact destabilizing naive multi-character tracking.

Correctness

The central claim, more stable tracking than baselines with more efficient training, rests on comparisons against unspecified baselines on InterHuman plus perturbation robustness tests per the introduction; the actual numbers sit in sections not read in this pass. Physics-based reinforcement learning trackers are also generally evaluated in simulation, so how the controller transfers to a production rig or to interactions outside boxing, pushing, and grappling is not addressed in what was read.

Clarity

The introduction is dense with prior-work citations but the core insight, that contact-rich two-character interaction needs to model force transfer or it destabilizes, is explained clearly with a concrete high-five example. A reader unfamiliar with reinforcement learning motion tracking, progressive neural networks, mixture of experts, will need background to follow the method beyond the intuition given here.

How to read it

First pass, read the abstract, Figure 1, and the contribution bullets to get the problem, force transfer instability, and the fix, progressive mixture of experts. Second pass, read the two-character interaction synthesis related-work subsection to see exactly how this differs from Won et al. and Zhu et al., since that comparison is the paper's main claim to novelty. Third pass, if relevant to a production reinforcement learning pipeline, read the method section on sample-to-expert routing and the results section for which interaction types were tested and how stability was measured.