Reproduce “Generating Continual Human Motion in Diverse 3D Scenes” 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

Aymen Mir, Xavier Puig, Angjoo Kanazawa, Gerard Pons-Moll

A framework synthesizes animator-guided, continual human motion in arbitrary 3D scenes by training on scene-agnostic motion capture data and utilizing a goal-centric canonical coordinate frame. The method achieves better perceived realism and reduced foot-skating (0.91 cm/frame) compared to existing baselines, enabling seamless transitions between diverse actions like walking, sitting, and grabbing.

Key insights

  • By training on scene-agnostic motion capture data and utilizing a goal-centric canonical coordinate frame, human motion can generalize effectively to diverse and arbitrary 3D scenes without explicit scene geometry conditioning.
  • Iteratively aligning the current motion with the next immediate goal in a canonical frame enables stable, long-range motion generation without accumulating drift.
  • Sparse action keypoints provide an intuitive and efficient interface for animators, allowing specification of complex, chained actions and interactions with minimal input.

Summary by alphaXiv.

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