RoboCasa is a large-scale simulation framework designed to train generalist robots in everyday environments, particularly in kitchen settings. The framework features diverse and realistic scenes, including 120 kitchen layouts and over 2,500 3D objects across 150 categories, created using generative AI tools. It supports cross-embodiment support for mobile manipulators and humanoid robots, and includes 100 diverse tasks, both atomic and composite, generated with the guidance of large language models (LLMs). RoboCasa provides a massive training dataset with over 100K trajectories, which is enriched with high-quality human demonstrations and automated trajectory generation methods. Experiments show that synthetically generated data in simulation can significantly improve policy learning and generalization, demonstrating the potential of simulation in scaling robot learning. The framework also explores the transfer of learned skills from simulation to real-world tasks, showing promising results.RoboCasa is a large-scale simulation framework designed to train generalist robots in everyday environments, particularly in kitchen settings. The framework features diverse and realistic scenes, including 120 kitchen layouts and over 2,500 3D objects across 150 categories, created using generative AI tools. It supports cross-embodiment support for mobile manipulators and humanoid robots, and includes 100 diverse tasks, both atomic and composite, generated with the guidance of large language models (LLMs). RoboCasa provides a massive training dataset with over 100K trajectories, which is enriched with high-quality human demonstrations and automated trajectory generation methods. Experiments show that synthetically generated data in simulation can significantly improve policy learning and generalization, demonstrating the potential of simulation in scaling robot learning. The framework also explores the transfer of learned skills from simulation to real-world tasks, showing promising results.