RoboCasa
SoftwareLarge-scale MuJoCo/robosuite household kitchen simulation framework, benchmark, and 2,200+-hour manipulation demonstration dataset (RoboCasa365) under MIT code and CC BY 4.0 asset/dataset licensing. [2] [3]
Installation & first result →Official websiteSource codeGetting started
- For
- Generalist robot policy researchers; Simulation benchmark developers; Imitation-learning and hierarchical-policy engineers [2]
- Hardware
- Simulated mobile manipulators (single-arm mobile platforms, humanoids, quadrupeds with arms), Optional keyboard or… [2]
- Source rights
- Named licenses are documented by part. Review licenses
Source researched · Not tested by us
Sources checked · Research details
Licenses & related hardware
- License
- MIT, CC-BY-4.0 · Details by part
Overview
RoboCasa (expanded in February 2026 as RoboCasa365) is a MuJoCo- and robosuite-backed household simulation framework and manipulation dataset spanning 365 tasks, 2,500+ kitchen scenes, 3,200+ 3D objects (human-designed/Objaverse and AI-generated), and 2,200+ hours of demonstrations (600+ hours human, 1,600+ hours synthetically generated). Recent 2026 updates added per-frame subtask annotations to the target composite task datasets, released a subtask annotation pipeline with labelers covering all 300 composite tasks, and launched a public leaderboard supporting generalist policies and agentic harnesses. [1] [2]
Is this for you?
- Workflows
- Gymnasium rollout and video capture
- Interactive kitchen scene, object, and demo playback
- Keyboard/SpaceMouse teleoperation
- Subtask annotation generation across 300 composite tasks
- Known limitations
- Requires cloning and installing ARISE-Initiative/robosuite from its master branch rather than a pinned PyPI release.
- setup.py enforces exact version pins (numpy==2.2.5, numba==0.61.2, scipy==1.15.3, mujoco==3.3.1, tianshou==0.4.10, lerobot==0.3.3), which can conflict with shared ML environments.
- Initial setup requires downloading ~10 GB of kitchen assets before running environments; v1.0.1 changed task horizon lengths by 1.5x for eval consistency.
What you need
- Operating system
- Linux
- macOS (use mjpython for MuJoCo viewer scripts)
- Windows
README states RoboCasa works across all major computing platforms and notes mjpython for macOS demos.
[2] · Editorial assessment
- Languages / runtimes
- Python
- Dependencies
- ARISE-Initiative/robosuite (master branch)
- mujoco==3.3.1
- numpy==2.2.5
- numba==0.61.2
- scipy==1.15.3
- tianshou==0.4.10
- lerobot==0.3.3
- gymnasium
- h5py
- Host computer
Requires ~10 GB disk space for kitchen assets download alone; demonstration dataset storage and GPU requirements depend on selected task splits and policy training stack.
- Compute for the workload
- ~10 GB disk space for kitchen assets download (excluding demonstration datasets)
Installation & first result
Open the upstream getting-started guide [2] · Project documentation [2]
- Create a Python 3.11 conda environment (conda create -c conda-forge -n robocasa python=3.11).
- Clone ARISE-Initiative/robosuite (master branch) and install editable with pip install -e .
- Clone robocasa/robocasa, install editable with pip install -e ., run python -m robocasa.scripts.setup_macros, and download ~10 GB of assets with python -m robocasa.scripts.download_kitchen_assets.
- Verify with gymnasium.make("robocasa/PickPlaceCounterToCabinet", split="pretrain") or demo_tasks (using mjpython on macOS).
- Installation
Create Python 3.11 conda env, install ARISE-Initiative/robosuite master branch editable, install robocasa editable, run setup_macros, and download ~10 GB kitchen assets.
Hardware, models & data
- Supported hardware
- Simulated mobile manipulators (single-arm mobile platforms, humanoids, quadrupeds with arms)
- Optional keyboard or SpaceMouse for teleoperation
- Interfaces
- Gymnasium API (gym.make("robocasa/..."))
- CLI demo/teleop/setup scripts (python -m robocasa.*)
- Keyboard and 3Dconnexion SpaceMouse teleoperation
- LeRobot / HDF5 dataset compatibility
- Inputs and outputs
Gymnasium-wrapped kitchen manipulation environments with pretrain/target splits, articulated appliances, and demonstration datasets with per-frame subtask annotations (subtask index, atomic skill, stage, language instruction).
- File formats
- HDF5 (h5py)
- LeRobot (lerobot==0.3.3)
- MuJoCo XML / 3D kitchen & Objaverse/AI-generated object assets
- Algorithms and models
- Benchmarking support for Diffusion Policy, pi, and GR00T (external policy models)
- Evaluation
Public RoboCasa365 Leaderboard comparing generalist policies (including Diffusion Policy, pi, GR00T, and agentic submissions); requires v1.0.1+ for updated 1.5x horizon lengths. Not executed in this review.
Project notes
Setup requires robosuite master, strict pins, and ~10 GB of assets
Installation is more tightly coupled than a standalone pip install: [2] [4]
1. Create and activate a Python 3.11 conda environment (conda create -c conda-forge -n robocasa python=3.11 followed by conda activate robocasa). 2. Clone ARISE-Initiative/robosuite on its master branch (specifically flagged as required in the README) and install it editable. 3. Clone robocasa/robocasa and install editable. Note that setup.py (version 1.0.1) enforces exact pins including numpy==2.2.5, numba==0.61.2, scipy==1.15.3, mujoco==3.3.1, tianshou==0.4.10, and lerobot==0.3.3—so isolate RoboCasa in its own environment to avoid breaking existing ML stacks. There is also an unresolved upstream conflict between the README's optional troubleshooting workaround (conda install -c numba numba=0.56.4 -y) and setup.py's numba==0.61.2 pin; dependency resolution was not tested in this review. 4. Run python -m robocasa.scripts.setup_macros followed by python -m robocasa.scripts.download_kitchen_assets, which pulls roughly 10 GB of kitchen assets before any task can run. That ~10 GB figure covers kitchen assets only; total download sizes for the 2,200+ hours of demonstration datasets were not established from the inspected README.
On macOS, interactive MuJoCo playback scripts (demo_tasks, demo_kitchen_scenes, demo_objects, demo_teleop) must be invoked with mjpython instead of python. If using a SpaceMouse for teleoperation, upstream notes you may need to configure SPACEMOUSE_PRODUCT_ID in robocasa/macros_private.py. [2]
Match v1.0.1+ before comparing leaderboard numbers
Environments are exposed through standard Gymnasium calls (gym.make("robocasa/PickPlaceCounterToCabinet", split="pretrain")) with explicit pretrain and target kitchen/object splits. If you plan to compare results against the RoboCasa365 Leaderboard, confirm you are on v1.0.1 or newer: the May 2026 v1.0.1 update increased horizon lengths by 1.5× across all tasks for evaluation consistency, so numbers from earlier v1.0 checkouts are not directly comparable. [2]
Separate MIT code, Apache-2.0 MuJoCo notice, and CC BY 4.0 data
The README splits licensing cleanly: code is under the MIT License, while 3D kitchen assets and demonstration datasets are under CC BY 4.0. In addition, the root LICENSE file appends an Apache-2.0 notice because the repository includes a partial implementation of DeepMind MuJoCo. Because the object library incorporates Objaverse and AI-generated assets, we recommend verifying third-party attribution notices before redistributing asset packs. [2] [3]
No conda environments were built, no 10 GB asset packs or demonstration datasets were downloaded, and no MuJoCo rollouts were executed in this review. The practical first step is creating an isolated Python 3.11 environment, pulling the kitchen assets, and running a single gym.make random rollout before downloading large demonstration splits. [2]
What we checked
Primary-source repository metadata, README, LICENSE, and setup.py inspection only. No conda environments were created, no 10 GB assets or demonstration datasets were downloaded, and no MuJoCo simulations were run.
Updates
Source files and licenses
Distinct large-scale household manipulation dataset and simulation benchmark with verified MIT code (plus Apache-2.0 MuJoCo notice) and CC BY 4.0 asset/dataset licensing.
| Layer | Availability / license |
|---|---|
| software | available · MIT Root LICENSE is MIT and includes an Apache-2.0 notice for partial DeepMind MuJoCo implementation. [2] [3] |
| datasets | available · CC-BY-4.0 README License section declares CC BY 4.0 for Assets and Datasets; Objaverse and third-party derived asset notices should be checked before redistribution. [2] |
| documentation | available · MIT Repository README covered by root MIT LICENSE; external docs portal hosted at robocasa.ai/docs/. [2] [3] |
RoboCasa provides simulation environments, assets, and demonstration datasets; external benchmarked policies (Diffusion Policy, pi, GR00T) are maintained separately.
References
Each source links to the exact revision we read. “Used for” shows where this page relies on it.
- S001 robocasa/robocasa repository metadata, v1.0 release, and HEAD commit
Confirms active 2026 maintenance, v1.0 RoboCasa365 release, and 2026-11-18 HEAD commit.Revision, date & where cited
RoboCasa Team · github.com · revision
7bfc60c· main HEAD (2026-11-18); release v1.0 (2026-02-18) · Accessed Nov 21, 2026
Used for: Overview, What we checked - S002 robocasa/robocasa README.md
Documents RoboCasa365 scope (365 tasks, 2,500+ kitchen scenes, 3,200+ 3D objects, 600+ hrs human + 1,600+ hrs synthetic demos), installation via robosuite master branch, ~10 GB kitchen assets download, Gymnasium API, and MIT/CC BY 4.0 licensing.Revision, date & where cited
RoboCasa Team · github.com · revision
7bfc60c· README.md at · Accessed Nov 21, 2026
Used for: Overview, Is this for you?, What you need, Installation & first result and 5 more - S003 robocasa/robocasa LICENSE
MIT License for RoboCasa code plus Apache-2.0 notice for included partial implementation of DeepMind MuJoCo.Revision, date & where cited
RoboCasa Team · github.com · revision
7bfc60c· LICENSE at · Accessed Nov 21, 2026
Used for: Separate MIT code, Apache-2.0 MuJoCo notice, and CC BY 4.0 data, Source files and licenses - S004 robocasa/robocasa setup.py
Declares version 1.0.1 and strict dependency pins including numpy==2.2.5, numba==0.61.2, scipy==1.15.3, mujoco==3.3.1, tianshou==0.4.10, lerobot==0.3.3, and gymnasium.Revision, date & where cited
RoboCasa Team · github.com · revision
7bfc60c· setup.py at · Accessed Nov 21, 2026
Used for: Is this for you?, What you need, Hardware, models & data, Setup requires robosuite master, strict pins, and ~10 GB of assets
Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1