Reproduce “Embodiment Meets Environment: Toward Context-Aware, Safe Physical Caregiving Robots” on OpenResearch

About this paper

Zhanxin Wu, Ruofei Tong, Jiaying Fang, Tapomayukh Bhattacharjee

E2-CARE is a framework that enables physical caregiving robots to achieve context-aware, safe adaptation across diverse environments and robot embodiments by dynamically synthesizing and enforcing interaction constraints from a unified 3D scene graph. It achieves high task success and constraint satisfaction rates in simulations and is perceived as significantly safer and more appropriate in real-world human-robot interaction studies.

Key insights

• Explicitly modeling the joint interaction between robot embodiment (and tools), the environment, and humans via a unified scene graph is fundamental for achieving context-aware safety and adaptation.
• Treating primitive skills as adaptable "interaction templates" that are dynamically reshaped by runtime constraints enables zero-shot reuse across diverse environments and robot embodiments without retraining.
• Utilizing control barrier functions allows for certified safety guarantees while adaptively modifying skill execution to satisfy task preferences, offering a robust approach in safety-critical, human-centered robotics.

Summary by alphaXiv.