SceneBot: New framework unifies humanoid free-space and contact-rich motion

SceneBot, developed by researchers at Stanford University (Sirui Chen, Shibo Zhao, Zhen Wu, Jiaman Li, Guanya Shi, C. Karen Liu), addresses a key limitation in humanoid robotics: existing reinforcement learning policies excel in free-space motions but fail at contact-rich tasks (e.g., carrying objects, traversing uneven terrain) because pure kinematic tracking cannot resolve the physical ambiguities of interaction. SceneBot overcomes this by conditioning a single policy on both reference motions and per-link contact labels, which explicitly define expected environmental interactions. To generate the necessary training data, the team created a hindsight scene reconstruction approach that infers scene-interaction graphs from retargeted human motion, providing 7.5 hours of reconstructed, contact-rich data.

Trained on this data, SceneBot demonstrates the ability to seamlessly unify free-space and contact-rich behaviors, executing complex, long-horizon tasks such as carrying a box upstairs for the first time in a general framework. Results show generalization to unseen motions and environments, establishing contact conditioning as a powerful interface for humanoid control. The paper (15 pages, 10 figures) is available on arXiv and all code and data will be open-sourced, with demos at the project website. This work significantly advances the field by bridging the gap between locomotion and manipulation in humanoid robots.