This paper introduces Moving Out, a novel human-AI collaboration benchmark that simulates diverse collaboration modes influenced by physical properties and constraints. It assesses the ability of agents to adapt to diverse human behaviors and unseen physical properties, including tasks such as co-moving heavy objects and navigating corners. Furthermore, to address the challenges of physical environments, we propose BASS (Behavior Augmentation, Simulation, and Selection), a novel method that enhances agent diversity and understanding of behavioral outcomes. We experimentally demonstrate that it outperforms state-of-the-art models in both AI-AI and human-AI collaboration.
