BEHAVIOR-1K

BEHAVIOR-1K is a large-scale simulation benchmark for human-centered embodied AI, developed by Stanford University (Chengshu Li, Ruohan Zhang, Josiah Wong, et al.), published in 2024. It was motivated by an extensive survey asking "what do you want robots to do for you?", resulting in 1,000 grounded everyday activities. The benchmark provides 50 richly detailed scenes including houses, gardens, restaurants, offices, and other indoor environments, populated with over 9,000 objects annotated with physical and semantic properties. These objects span rigid bodies, deformable bodies, and liquids, enabling realistic simulation of diverse household tasks. The simulation environment, OmniGibson, supports realistic physics simulation and rendering of all object types. Activities in BEHAVIOR-1K are characteristically long-horizon and dependent on complex manipulation skills — representing a significant challenge for even state-of-the-art robot learning solutions. BEHAVIOR-1K represents a shift from traditional robot benchmarks focused on isolated pick-and-place or navigation tasks, toward human-centric activities that reflect real household needs. It includes kitchen tasks (cooking, cleaning), laundry, gardening, pet care, and many other everyday activities. The benchmark includes an initial simulation-to-reality transfer study with a mobile manipulator, calibrating the gap between simulation and real-world deployment. It serves as a critical resource for developing and evaluating generalist robot policies capable of handling diverse, long-horizon manipulation tasks.