NTU's MAMHOI factorizes scene-aware human-object interaction via affordances
nanyang-technological-university-singapore · hf · 2026-10-12
Researchers at Nanyang Technological University present MAMHOI, an affordance-mediated factorization for scene-aware human-object interaction (HOI) generation in complex 3D scenes.
- Problem: realistic HOI requires both environment feasibility reasoning and realistic human-object motion synthesis, but paired human-object-scene supervision is scarce while human-scene and human-object datasets each cover only one side.
- Method: an explicit motion-affordance interface sits between scene understanding and motion synthesis — a scene-conditioned model first predicts where and how an interaction can feasibly be executed, then an affordance-conditioned HOI model generates the corresponding human-object motion, letting each capability learn from complementary supervision without paired data.
- Results: in complex indoor environments, MAMHOI reduces object–scene penetration while better preserving interaction quality, yielding more physically feasible interactions.
Project page: https://leimingyuan.github.io/MAMHOI-project-page/
More from Embodied
- The real2sim2real^2 meme: a wall of robot sim training monitors — chris_j_paxton · 2026-10-12
- ByteDance world model demoed in five use cases, from film camerawork to robot training — oran_ge · 2026-10-12
- AGIBOT G2 humanoid assists nurses in operating rooms at Fudan-affiliated hospital — CyberRobooo · 2026-10-12
- Dev uses Codex to write build tutorials and code for his MOSS robot — OpenAIDevs · 2026-10-12
- Amazon now runs 1 million+ robots alongside 1.5 million human workers — Ars Technica AI · 2026-10-12
- DeepSeek reportedly leads tens of millions seed round for robotics startup Winverse AI at ~$400M valuation — zephyr_z9 · 2026-10-12