RoboCasa – Large‑Scale Robot Simulation Benchmark

What it is

RoboCasa (now RoboCasa365) is an open‑source simulation framework that lets researchers train and evaluate robot policies on a huge variety of everyday kitchen tasks. It builds on the robosuite physics engine and provides:

365 distinct tasks (e.g., pick‑place, opening cabinets) generated with the help of large language models.
2,500+ kitchen scenes and 3,200+ 3‑D objects, including both human‑designed and AI‑generated items.
Hundreds of hours of demonstration data – human‑recorded demos plus robot‑generated trajectories, now enriched with per‑frame sub‑task annotations.
Benchmarking tools: ready‑made Gym environments, a leaderboard, and scripts for training popular policies such as Diffusion Policy, PI, and GR00T.

In short, it is a research‑grade “playground” for building generalist robot agents that can handle many household manipulation tasks in simulation.

How to get started

Create a conda environment (Python 3.11) and install the required robosuite backend from its master branch.
Clone RoboCasa, install it in editable mode, and optionally set up the pre‑commit formatter.
Download assets – a ~10 GB payload containing the kitchen scenes and object models.
Run a quick demo – the repository ships with scripts to:
- launch a Gym environment and record random rollouts,
- play back example demonstrations,
- browse the kitchen scenes or object library,
- tele‑operate the robot via keyboard or a SpaceMouse.

All of this can be done with a few Python commands as shown in the README.

What you can do with it

Use‑case	What RoboCasa provides
Policy research	Gym‑compatible environments, pre‑packaged task definitions, and a large, labeled demonstration dataset for imitation or reinforcement learning.
Hierarchical learning	Per‑frame sub‑task labels (skill name, stage, natural‑language instruction) enable training of hierarchical or skill‑based policies.
Benchmarking	Standardized evaluation splits (`pretrain` vs `target` kitchens), a public leaderboard, and reference implementations of several state‑of‑the‑art algorithms.
Data‑driven asset creation	Access to both curated and AI‑generated objects (via Objaverse) for testing generalization to novel items.
Simulation research	Photo‑realistic rendering (via RoboSuite v1.5) and flexible robot composition for studying controller design.

Who might find it useful?

Robotics researchers building manipulation policies that need to generalize across many tasks and objects.
Machine‑learning scientists exploring large‑scale imitation learning, diffusion policies, or hierarchical RL.
Students looking for a ready‑made benchmark to test new algorithms without building a simulator from scratch.
Community contributors who can submit their own models to the leaderboard and compare against published baselines.

Limitations & considerations

Simulation‑only: The framework runs in the RoboSuite physics engine; transferring policies to real robots will require additional sim‑to‑real work.
Asset size: Downloading the full dataset (~10 GB) is required for most tasks, which may be a barrier on limited bandwidth or storage.
Hardware: Running the photo‑realistic rendering and large‑scale training benefits from a GPU and a decent CPU; pure‑CPU setups will be slower.
Scope: Focused on kitchen manipulation; other domains (e.g., mobile navigation) are not covered out of the box.

Quick code snippet

import gymnasium as gym
import robocasa
from robocasa.utils.env_utils import run_random_rollouts

env = gym.make(
    "robocasa/PickPlaceCounterToCabinet",
    split="pretrain",
    seed=0,
)
run_random_rollouts(env, num_rollouts=2, num_steps=120, video_path="demo.mp4")

This creates a pick‑and‑place task in a randomly chosen kitchen scene and records a short video of random actions.

Where to learn more

Home page & docs: https://robocasa.ai
Paper (ICLR 2026): https://robocasa.ai/assets/robocasa365_iclr26.pdf
Original RSS 2024 paper: https://robocasa.ai/assets/robocasa_rss24.pdf
Leaderboard: https://robocasa.ai/leaderboard.html

License & citation

Code is released under the MIT License.
Assets and datasets are under CC‑BY 4.0.
Cite the appropriate paper (RoboCasa365 or the original RoboCasa) when using the framework or data.

robocasa/robocasa