RoboCasa365 — Large-Scale Simulation of Everyday Tasks for Generalist Robots

TL;DR

RoboCasa is a large-scale kitchen-domain simulation framework for training generalist robot manipulation policies, originally released in 2024 (RSS) and re-released as RoboCasa365 (ICLR 2026) with significant scale-ups. The v1.0 drop ships 365 LLM-guided tasks across 2,500+ kitchen scenes and 3,200+ 3D objects, ~600 hours of human demonstrations, and 1,600+ hours of robot trajectories generated by automated tooling — plus baseline integrations with Diffusion Policy, π (pi0), and GR00T and a public leaderboard. Built on robosuite v1.5, it’s MIT-licensed (code) and CC BY 4.0 (assets/datasets), pip-installable in a conda env, and supports keyboard / SpaceMouse teleop. For Luma it’s the closest thing in the open ecosystem to a standardized embodied-AI eval substrate at task scale — the manipulation analog of OpenReward / Toolathlon-GYM for terminal agents.

Key claims

365 manipulation tasks created with LLM guidance, plus 2,500+ kitchen scenes and 3,200+ objects (both human-designed and AI-generated), all packaged as a single pip-installable robocasa package on top of robosuite v1.5 [README §Overview, §Releases v1.0].
The dataset bundles ~600 hours of human teleop demonstrations and 1,600+ hours of robot trajectories generated by “automated trajectory tools” — i.e. the demo set itself is a hybrid of human and auto-synthesized supervision, not pure teleop [README §Overview].
Benchmarking ships with three reference policy-learning baselines (Diffusion Policy, π / pi0, GR00T) and a hosted leaderboard, framing RoboCasa as an evaluation platform first and a training platform second [README §Overview, §Tasks/datasets/policy learning].
Gym wrapper is the primary entry point: gym.make("robocasa/PickPlaceCounterToCabinet", split="pretrain"|"target", seed=...) with a split parameter that explicitly separates pretrain and target kitchen scenes / objects — i.e. generalization to held-out scenes is a first-class evaluation axis [README §Gym wrapper].
v1.0.1 (May 2026) increased horizon lengths 1.5× across all tasks for consistency; v1.0 (Feb 2026) is the RoboCasa365 release; v0.2 (Oct 2024) was the move to robosuite v1.5 with photo-realistic rendering [README §Releases].
Asset download is ~10 GB and is a separate download_kitchen_assets step from the pip install — the simulation framework and the asset library are decoupled [README §Installation step 5].

Method

RoboCasa is built as a thin task-and-asset layer on top of robosuite (the underlying MuJoCo-based manipulation simulator from ARISE-Initiative). The framework provides three orthogonal axes of scale: (a) tasks — 365 manipulation problems with LLM-assisted authoring; (b) scenes — 2,500+ procedurally varied kitchen layouts; (c) objects — 3,200+ assets, including both Objaverse-sourced human-designed models and AI-generated models (toggleable via --obj_types aigen). The standard Gymnasium API exposes each (task, split, seed) triple as a callable environment, with split=pretrain|target partitioning scenes and objects to test cross-scene generalization. For training, the demo bundle pairs 600+ hours of human teleop (keyboard / SpaceMouse, with the robot rendered semi-translucent to reduce occlusion) with 1,600+ hours of automated-trajectory-tool output — the latter is how the 365-task × 2,500-scene matrix gets covered without proportional human cost. Three reference policies (Diffusion Policy, pi0, GR00T) are wired up as baselines, and a public leaderboard accepts user-submitted models.

Results

The README is a release artifact, not a paper, so headline numbers are not in scope here — they live in the linked RoboCasa365 ICLR 2026 paper and the RSS 2024 original. What the release does commit to quantitatively: 365 tasks, 2,500+ scenes, 3,200+ objects, 600+ hours human demos, 1,600+ hours robot demos, ~10 GB asset download, and three reference baselines on a hosted leaderboard. The 1.5× horizon increase in v1.0.1 is flagged as a breaking change for eval consistency.

Why it’s interesting

RoboCasa is the manipulation-domain answer to the “env-as-package” pattern that RL Environment Platforms traces for terminal/coding agents — same shape (Gym-wrapped tasks, conda-installable, automated verifier per task, leaderboard) but ported to embodied control with a real physics engine instead of a Linux sandbox. It complements Genesis World — Simulation platform for general-purpose robotics & embodied AI learning (Genesis World) and Project Chrono — Open-Source Multi-Physics Simulation Engine (Project Chrono) on the simulator-platform axis, but unlike those it ships with a curated 365-task benchmark and a pre-recorded 2,200+-hour demo set, making it more directly comparable to EgoVerse: An Egocentric Human Dataset for Robot Learning from Around the World and 1X World Model (1XWM) — action-controllable world model for humanoid policy evaluation on the supervised-trajectory side. The hybrid 600h-human + 1,600h-auto demo split is notable: it’s the same “synthesizer-driven task generation” pattern SETA: Scaling Environments for Terminal Agents (env repo) uses in the terminal domain (frontier model authors tasks, smaller model trains on them) ported to robot trajectories. RoboCasa365 also predates and motivates several pieces in the recent robotics-eval-via-video-models thread (Evaluating Gemini Robotics Policies in a Veo World Simulator, Dream.exe: Can Video Generation Models Dream Executable Robot Manipulation?) — those papers use video models to evaluate policies that are typically trained on RoboCasa-like substrates.