RoboCasa
New1. Overview
RoboCasa is a large-scale simulation framework for training and benchmarking generalist robots on everyday household tasks, developed by researchers at UT Austin with support from NVIDIA Research and originally released in 2024 at RSS. Its focus is kitchen environments, which it models from real-world layouts and styles and fits with interactable furniture and appliances. The current release, RoboCasa365, was published at ICLR 2026 and expands the platform into a comprehensive benchmark for household mobile manipulation: 365 everyday tasks across 2,500 diverse kitchen environments, with more than 600 hours of human demonstration data and more than 1,600 hours of synthetically generated demonstrations. The design targets three problem settings explicitly: multi-task learning, robot foundation model training and lifelong learning. For a physics-AI programme the significance is that it turns the environment side of embodied AI into a reproducible, scalable asset: instead of arguing about how close a policy is to general capability, a team can run the same task suite at the same scale and compare numbers on a public leaderboard.
2. How It Works
The stack builds on robosuite with MuJoCo as the physics backend, so scenes are articulated rigid-body models with contact, mass and friction rather than purely visual proxies. Scenes are authored from kitchen layouts compiled by consulting architecture and home design references, modelled to standard size and spatial specifications, then populated with a repository of interactable furniture and appliances spanning cabinets, stoves, sinks and microwaves. Many of these are articulated, so doors open, knobs twist and burners switch on when the corresponding control is actuated, which means the manipulation problem includes state changes in the environment rather than only object transport. Task creation is LLM-guided: GPT-4 proposes task candidates conditioned on the available environments and the robot's skill vocabulary, and the resulting task set is split into 65 atomic tasks covering ten foundational skills and 300 composite tasks that sequence those skills into semantically meaningful activities such as restocking supplies or brewing coffee. A subtask annotation pipeline labels every frame of a demonstration with a subtask index, atomic-skill name, stage and natural-language instruction, which supports hierarchical policy learning. Scene appearance is varied with AI-generated textures for walls, floors, counters and cabinet panels, and object assets are curated from Objaverse 1.0, LightWheel AI and Luma AI generation.
3. Scope and Assumptions
The framework targets household mobile manipulation, specifically kitchen scenes, and the benchmark is built around ten foundational skills: pick and place, opening and closing doors, opening and closing drawers, twisting knobs, turning levers, pressing buttons, insertion, navigation, sliding racks, and opening and closing lids. It supports mobile manipulators of several form factors, including single-arm mobile platforms, humanoid robots, and quadruped robots with arms, so results are not tied to one embodiment. It is a simulation benchmark, not a deployment stack: policies are trained and evaluated in simulation, and any real-world claim requires a separate transfer study. It assumes the MuJoCo and robosuite dependency chain is installed correctly and that roughly ten gigabytes of kitchen assets have been downloaded through the provided script. The task set is kitchen-centric by construction, so conclusions about general household capability are bounded by that choice, and the object library inherits the licence terms of its upstream sources, which vary per asset.
4. Getting Started
The documented path is a conda environment on Python 3.11, then robosuite installed from its master branch, then this repository installed in editable mode, followed by two setup scripts that configure system variables and download the kitchen assets, which the README flags at about ten gigabytes. Environments are created through a gymnasium wrapper with a task name such as PickPlaceCounterToCabinet, a split argument that selects either the pretrain or the target kitchen scenes and objects, and a seed for reproducibility. The repository ships demo entry points for playing back sample task demonstrations, exploring the kitchen scenes, browsing the object library, and teleoperating the robot with a keyboard or a SpaceMouse, which is the fastest way to understand what a task actually requires before training anything. A sensible first experiment is to run a random rollout with video recording, then play back a human demonstration of the same task and compare the action space and horizon, and only then set up the evaluation harness, because the task horizon changed substantially in version 1.0.1 and old numbers are not comparable with new ones.
5. Comparison with Alternatives
Against raw physics engines such as MuJoCo or GPU-resident solvers such as Newton, RoboCasa is an application-level benchmark rather than an engine, and it consumes those engines instead of competing with them. Against motion and whole-body frameworks such as ProtoMotions and SuperMimic, the difference is task domain and evaluation contract: those concentrate on locomotion and motion tracking, while RoboCasa concentrates on long-horizon household manipulation with a fixed task suite and a leaderboard. Against asset collections, the relationship is complementary rather than parallel: the RoboCasa365 scene and object entries index the assets under a simulation-readiness grade, while this entry covers the framework, the task suite and the benchmark protocol, and comparable asset libraries such as MuJoCo Menagerie and Open X-Embodiment supply models and real-robot data respectively. What distinguishes RoboCasa in practice is the combination of LLM-authored tasks, articulated scene state and a public leaderboard, which is closer to a benchmark contract than to a simulator feature list.
6. Benchmarks and Evidence
The RoboCasa365 paper reports a comprehensive benchmark for household mobile manipulation and runs extensive experiments with state-of-the-art methods, analysing how task diversity, dataset scale and environment variation affect generalization. The headline resources are 365 tasks across 2,500 kitchen scenes, more than 600 hours of human demonstrations and more than 1,600 hours of synthetic demonstrations. The benchmark supports multi-task learning, foundation model training and lifelong learning, and the leaderboard compares generalist policies including Diffusion Policy, pi-zero and GR00T, while also accepting community submissions and, since October 2026, agentic submissions built around a base policy. Version history matters for reproducibility: the 1.0 release landed in February 2026 with 365 tasks, 2,500 scenes and 2,200-plus hours of demonstration data, and version 1.0.1 in May 2026 increased task horizons by about fifty percent for consistency, which means evaluation numbers from before and after that change are not directly comparable. The paper's findings about which factors most strongly affect generalist performance are the authors' own analysis on their benchmark rather than an independent third-party replication.
7. Role in Physics-AI Workflows
RoboCasa occupies the environment layer of an embodied-AI pipeline: it supplies the scenes, the tasks and the evaluation contract that a policy is trained and measured against. Its contribution to a physics-AI programme is less about new physics and more about scale and comparability, because the hard part of generalist robot learning is not a single simulation but a reproducible suite that reveals whether a method generalizes across task diversity and environment variation. The articulated scene state is where physics matters: doors, drawers, knobs and burners are joints and state machines, so a policy that only matches pixels fails, and contact, friction and mass properties propagate into every demonstration. Its synthetic-data pillar connects to the broader data question covered by Open X-Embodiment and to embodied platforms indexed elsewhere as AgiBot World and ALOHA. The physics simulation primer covers the modelling assumptions behind articulated scenes, Python robotics covers the tooling that connects a policy to a controller, and the physics-AI learning path places this environment stage alongside the asset side represented by Isaac Sim assets.
8. Limitations and Failure Modes
The first limitation is domain scope: the suite is kitchen-centric, so a policy that scores well here has demonstrated kitchen household competence, not general household competence, and claims should be phrased accordingly. The second is sim-to-real distance: everything is evaluated in MuJoCo, and textures and object variety are synthetic, so visual and dynamic gaps with a physical kitchen are inherited by every result. The third is version drift: the task horizons changed by roughly fifty percent in version 1.0.1 and subtask annotations were added mid-2026, so numbers from different releases must not be mixed, and a comparison without a pinned version is meaningless. The fourth is asset licensing: the object library draws on Objaverse 1.0, LightWheel AI and Luma AI generation, so commercial use requires per-asset verification even though the framework code itself is MIT. The fifth is the compute and storage footprint, with roughly ten gigabytes of assets and long-horizon tasks that make large-scale training expensive. The sixth is a documentation-level conflict worth flagging: the repository LICENSE declares MIT for the code and notes embedded Apache-2.0 MuJoCo portions, and the README declares CC BY 4.0 for assets and datasets, so any downstream licence statement must be checked against those two primary sources rather than against secondary summaries.
9. Version and Source Notes
RoboCasa was originally released in 2024 at RSS as "RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots" by Soroush Nasiriany, Abhiram Maddukuri, Lance Zhang, Adeet Parikh, Aaron Lo, Abhishek Joshi, Ajay Mandlekar and Yuke Zhu. The current benchmark, RoboCasa365, appears at ICLR 2026 as arXiv 2603.04356, submitted 4 March 2026, by Soroush Nasiriany, Sepehr Nasiriany, Abhiram Maddukuri and Yuke Zhu, with the first three authors contributing equally; the core team is at UT Austin with NVIDIA Research support. The repository LICENSE file declares MIT for the code with copyright held by the RoboCasa Team in 2026, and states that the software includes partial implementation of DeepMind MuJoCo, which remains under Apache-2.0; the README separately declares MIT for code and CC BY 4.0 for assets and datasets. Release history runs from version 0.2 in October 2024, which moved to robosuite 1.5 with composite controllers and photo-realistic rendering, through version 1.0 in February 2026 with 365 tasks, 2,500-plus scenes and 2,200-plus hours of demonstrations, version 1.0.1 in May 2026 with longer horizons, per-frame subtask annotations in July 2026, a subtask annotation pipeline covering all 300 composite tasks in October 2026, and agentic leaderboard submissions from 28 October 2026. This entry treats the repository LICENSE, the README and the two papers as primary sources and makes no inference about personnel or funding beyond what they state.
10. References
- RoboCasa project page, robocasa.ai, overview, environment gallery, update log and leaderboard.
- Nasiriany, Nasiriany, Maddukuri, Zhu, "RoboCasa365: A Large-Scale Simulation Framework for Training and Benchmarking Generalist Robots", ICLR 2026, arXiv 2603.04356.
- Nasiriany, Maddukuri, Zhang, Parikh, Lo, Joshi, Mandlekar, Zhu, "RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots", RSS 2024.
- RoboCasa repository, github.com/robocasa/robocasa, installation guide, demo scripts and license statement.
- RoboCasa documentation, robocasa.ai/docs, task catalogue, dataset descriptions and benchmarking guide.
- RoboCasa365 leaderboard, robocasa.ai/leaderboard.html, generalist policy comparison including community and agentic submissions.
- robosuite, the manipulation simulation framework that RoboCasa builds upon as its environment layer.
- MuJoCo, the physics engine underlying robosuite, whose Apache-2.0 portions are embedded per the repository LICENSE.
- Objaverse 1.0, one of the upstream sources of the object asset library, with per-asset license terms.
- RoboCasa365 scene and object entries on this site, the asset-side view of the same platform under a simulation-readiness grade.