Reproduce “HOSIG: Full-Body Human-Object-Scene Interaction Generation with Hierarchical Scene Perception” 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
Wei Yao, Yunlian Sun, Hongwen Zhang, Yebin Liu, Jinhui Tang
HOSIG, developed by a collaborative team including researchers from Tsinghua University, presents a framework for full-body human-object-scene interaction generation by integrating scene-aware navigation with dexterous object manipulation. The system achieved a 93.3% success rate in object locomotion tasks and reduced human-scene penetration rates by 42-55% compared to prior methods.
Key insights
- Integrating hierarchical scene perception (local for grasping, global for navigation, and path-aligned for motion refinement) is crucial for comprehensive and robust human-object-scene interaction generation.
- Decoupling the complex HOSI task into specialized modules for grasp generation, navigation planning, and motion synthesis, operating in an iterative pipeline, enhances both precision and coherence.
- Utilizing gradient-based scene guidance during diffusion inference, combined with a global 6D parameterization for hand poses, enables fine-grained control and collision avoidance without over-constraining the motion generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.