RoboCasa — Large-Scale Simulation for Everyday Household Tasks
RoboCasa is a large-scale simulation environment and dataset developed at UT Austin for training household manipulation policies. Released in 2024 under MIT license, it contains 100,000 demonstrations spanning 100 diverse kitchen and household tasks simulated with photorealistic assets in 150 scene variations. Built on the MuJoCo physics engine and the Robosuite framework, RoboCasa provides the most comprehensive simulated kitchen environment available, covering tasks from opening cabinets and operating appliances to food preparation and object organisation. The dataset is designed as a bridge between simulation and real-world deployment, with scene diversity intended to force policies to generalise rather than memorise specific environments. RoboCasa is particularly significant for the food service and household service automation research community as it provides standardised evaluation across 100 clearly-defined tasks with ground truth success metrics.
| Year | 2024 |
|---|---|
| Episodes | 100,000 |
| Embodiments | Franka Panda (simulated) |
| Modalities | rgb, depth |
| Task categories | manipulation, cooking, cleaning, pick-and-place, long-horizon |
| Data format | json, mp4 |
| License | MIT |
| Access | open — commercial use permitted |
| Maintainer | UT Austin |
| Origin country | US |
What is it?
RoboCasa is a large-scale simulation environment and dataset developed at UT Austin for training household manipulation policies. Released in 2024 under MIT license, it contains 100,000 demonstrations spanning 100 diverse kitchen and household tasks in 150 scene variations, built on MuJoCo and Robosuite. It is the most comprehensive simulated kitchen environment available for robot learning research.
Who is it for?
Researchers studying household robot generalisation — whether policies trained in simulation can transfer to real kitchens. Particularly valuable for food service and domestic robot automation research, providing standardised evaluation across 100 clearly-defined tasks with ground truth success metrics.
Key specifications
- Demonstrations: 100,000 episodes
- Tasks: 100 distinct household and kitchen tasks
- Scene variations: 150 different environments
- Robot platform: Simulated Franka Panda
- Physics engine: MuJoCo via Robosuite
- Format: JSON, MP4
- License: MIT — commercial use permitted
- Access: Open — GitHub and project website
How it compares
RoboCasa is the largest structured kitchen simulation dataset by task count (100 tasks). LIBERO covers 130,000 episodes across 4 task suites but with less kitchen-specific depth. Meta-World covers 50 tasks but focuses on tabletop manipulation rather than household environments. RoboCasa's kitchen specificity and scene diversity make it the reference benchmark for household robot generalisation.
Limitations and access notes
Simulation-only — sim-to-real transfer remains an open challenge for kitchen environments. MIT license permits unrestricted commercial use.
Linked professions
Frequently asked questions
What is RoboCasa?
RoboCasa is a simulation dataset and benchmark covering 100 household and kitchen manipulation tasks across 150 scene variations. It contains 100,000 demonstration episodes for a simulated Franka Panda robot, designed to train and evaluate generalised household robot policies.
Can RoboCasa be used commercially?
Yes. RoboCasa is MIT licensed, permitting unrestricted commercial use, modification, and redistribution.
What kitchen tasks does RoboCasa cover?
RoboCasa covers 100 tasks including opening cabinets and drawers, operating appliances (microwave, oven, dishwasher), food preparation, object placement, and household organisation tasks across 150 scene variations including different kitchen layouts and visual conditions.
How does RoboCasa differ from real-robot datasets?
RoboCasa is purely simulation-based — no physical robot was used. This enables large-scale data collection at low cost but introduces a sim-to-real gap: policies must adapt to real-world physics, lighting, and object variations not present in simulation.
How do I access RoboCasa?
RoboCasa is available at robocasa.ai and via the GitHub repository github.com/robocasa/robocasa. No registration is required.