365 Tasks, 2,500 Kitchens, and 600 Hours of Demos: Inside RoboCasa

Training a robot to reliably unload a dishwasher or wipe down a counter is hard, and a big part of that difficulty is that you need data. Lots of it. You need varied environments, varied objects, and demonstrations of humans actually doing the tasks. RoboCasa is a simulation framework built to solve exactly that problem—generating the scale of training data that general-purpose robot policies need.

What It Does

RoboCasa is a large-scale simulation framework for training generally capable robots to perform everyday tasks. It started as a 2024 release from UT Austin researchers, and the latest iteration—RoboCasa365—expands it into a platform for large-scale training and benchmarking in simulation.

The framework is organized around four pillars. First, diverse tasks: 365 of them, created with the guidance of large language models. Second, diverse assets: over 2,500 kitchen scenes and 3,200 3D objects. Third, high-quality demonstrations, which include more than 600 hours of human demonstrations plus over 1,600 hours of robot datasets generated with automated trajectory tools. And fourth, benchmarking support, with implementations of popular policy learning methods like Diffusion Policy, pi, and GR00T, along with a public leaderboard for user-submitted models.

On the technical side, it builds on RoboSuite as its backend (the README specifically calls for the master branch) and works across all major computing platforms, with Conda as the recommended setup path.

Why It's Cool

The scale is the point. A single kitchen and a handful of objects won't get you a policy that generalizes. RoboCasa365 gives you 2,500+ kitchen scenes and 3,200+ objects to train against. That's the kind of variety that forces a model to learn something general rather than memorizing one layout.

The demonstration data is unusually honest about where it comes from. Rather than pretending all data is equal, the project separates 600+ hours of human demonstrations from 1,600+ hours of robot data created with automated trajectory tools. That distinction matters—you can decide how much to weight each source depending on what you're training.

The subtask annotations are a genuinely useful detail. As of July 2026, composite task datasets include per-frame annotations: every timestep is labeled with a subtask index, an atomic-skill name, a stage (pick, place, or navigate), and a natural-language instruction. If you're working on hierarchical policies—where a high-level planner picks a skill and a low-level controller executes it—this saves you from building that labeling pipeline yourself.

It doesn't lock you into one policy architecture. Diffusion Policy, pi, and GR00T are all supported out of the box, and the leaderboard means you can see how your approach stacks up against others without setting up your own comparison harness.

Version 1.0.1 quietly fixed something that trips up benchmarking. Horizon lengths were increased 1.5x across all tasks for consistency. If you'd been comparing eval numbers across tasks and something felt off, that's likely why—and it's worth updating before running evals.

How to Try It

The setup is Conda-based and straightforward. Here's the flow:

1. Create and activate the environment:

   conda create -c conda-forge -n robocasa python=3.11
   conda activate robocasa
   

2. Clone and install RoboSuite—use the master branch, this matters:

   git clone https://github.com/ARISE-Initiative/robosuite
   cd robosuite
   pip install -e .
   

3. Clone and install RoboCasa itself:

   cd ..
   git clone https://github.com/robocasa/robocasa
   cd robocasa
   pip install -e .
   pip install pre-commit; pre-commit install
   

   (There's an optional fix if you hit numba/numpy issues: conda install -c numba numba=0.56.4 -y.)

4. Set up system variables and download the assets:

   python -m robocasa.scripts.setup_macros
   python -m robocasa.scripts.download_kitchen_assets
   

Heads up on that last step: the assets are around 10GB. Once that's done, you can create environments through gym wrappers and run rollouts—the README shows a gymnasium import as the starting point.

Full docs live at robocasa.ai/docs, and the repository is at github.com/robocasa/robocasa.

Final Thoughts

RoboCasa365 is aimed squarely at researchers working on robot learning in simulation—people who need a lot of task variety and demonstration data without building a data collection pipeline from scratch. The 10GB asset download and the RoboSuite dependency mean it's not a five-minute install, but that's the tradeoff for getting thousands of scenes out of the box. If you're training or benchmarking manipulation policies and you've been bottlenecked on environment diversity, this is worth the setup time.