Reproduce “Learning Reactive Human Motion Generation from Paired Interaction Data Using Transformer-Based Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Masato Soga, Ryuki Takebayashi

Researchers developed Transformer-based models to generate reactive human motion in two-person interactive scenarios, specifically boxing, evaluating the effectiveness of a simple Transformer, iTransformer, and Crossformer architectures, along with a person ID embedding. The simple Transformer achieved the highest scores for motion quality, and person ID embedding consistently improved generated motion stability and human-likeness.

Key insights

  • The Simple Transformer, by processing all joints simultaneously via temporal self-attention, implicitly maintains global pose consistency and avoids the 'posture collapse' observed in other models.
  • Models that treat features (joints) as independent tokens or model dependencies in stages (iTransformer, Crossformer) can accumulate joint-wise errors, leading to unstable and implausible postures.
  • Person ID embedding significantly enhances structural consistency and stabilizes motion generation, demonstrating its importance for robust multi-person motion synthesis.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.