Moving Out: Physically-grounded Human-AI Collaboration
Researchers have introduced "Moving Out," a new benchmark designed to evaluate human-AI collaboration in physical environments. This benchmark addresses the limitations of existing discrete collaboration tasks by incorporating continuous state-action spaces and physical constraints. The goal is to enable embodied agents, such as robots, to effectively work alongside humans by adapting to physical actions and environmental dynamics. To achieve this, a novel method called BASS (Behavior Augmentation, Simulation, and Selection) was developed to improve agent diversity and their understanding of action outcomes, demonstrating successful collaboration with both unseen AI agents and humans. AI
IMPACT This benchmark could accelerate the development of more capable embodied AI agents for real-world human-AI collaboration.