RoboCasa

RoboCasa is a large-scale simulation framework developed by researchers at UT Austin and NVIDIA for training generalist household robots. Built on top of robosuite and MuJoCo, RoboCasa provides photorealistic kitchen and household environments with procedural variation for diverse robot learning tasks. The dataset includes 120+ distinct 3D scenes with over 10,000 unique objects organized across 20 task families. Each task family contains procedurally generated variations in object placement, scene layout, and visual appearance. Over 5,000 expert demonstration trajectories are included for imitation learning and reward design. RoboCasa supports research in sim-to-real transfer, generalizable manipulation, and household robotics. Its focus on everyday tasks (opening cabinets, picking up objects, cleaning, organizing) makes it a critical resource for developing robots that can operate in unstructured human environments.