RoboCasa: Kitchen Manipulation Dataset
Large-scale simulation data for household robot policies. 536K+ downloads. MIT licensed and free to use.
Key Stats
| Metric | Value |
|---|---|
| Size | ~20 GB (pretrain); multiple variants total 100+ GB |
| Downloads | 536K+ (one of the most downloaded robotics datasets) |
| Tasks | Kitchen and household manipulation (cooking, cleaning, organizing) |
| Format | Parquet (pretrain); Parquet + MP4 (LeRobot variants) |
| Modalities | RGB images, joint state, actions |
| License | MIT |
| Variants | 406 total RoboCasa datasets on HuggingFace |
What is RoboCasa?
RoboCasa is a large-scale simulation environment and dataset suite designed specifically for household robotics. It provides diverse, realistic kitchen environments populated with everyday objects and task scenarios covering cooking, cleaning, and organization. The pretraining dataset (robocasa-pretrain) has become one of the most downloaded robotics datasets on HuggingFace with over 536K downloads.
The RoboCasa ecosystem includes several variants: the base pretrain dataset, NVIDIA's RoboCasa-Cosmos-Policy (2.98M rows formatted for Cosmos world model training), and LeRobot-compatible versions (robocasa_target_human_unified with 15M rows). There are 406 total RoboCasa-related datasets on HuggingFace.
RoboCasa is particularly valuable as a pretraining corpus for kitchen manipulation policies that will later be fine-tuned on real-world data.
Related datasets
- ManiSkill2 -- another simulation manipulation benchmark
- LeRobot Collection -- RoboCasa variants available in LeRobot format
- NVIDIA GR00T Teleop -- NVIDIA's Cosmos integration with RoboCasa