Real-World Data for RoboCasa

RoboCasa simulates realistic kitchen tasks. Real-world data adds the material interactions, thermal effects, and environmental chaos of actual cooking.

RoboCasa at a Glance

100
Tasks
2,500+
3D Assets
120+
Kitchen Configs
150+
Object Categories
MuJoCo
Physics Engine
2024
Released

Kitchen Task Categories

RoboCasa's tasks span the full range of kitchen manipulation, each presenting different physical interaction challenges.

CategoryExample TasksKey ChallengeSim-to-Real Gap
Appliance OperationOpen fridge, microwave, ovenHandle detection, hinge dynamicsReal hinge resistance, magnetic latches
Item RetrievalGet item from shelf, fridgeSpatial planning, grasp in clutterObject weight variation, shelf friction
Surface CleaningWipe counter, clean stoveContact-rich wiping motionReal friction, material compliance, cleaning fluid
Food PreparationPour, stir, cutMaterial deformation, fluid dynamicsLiquid viscosity, food texture, knife friction
Multi-Step CookingMake coffee, prepare ingredientsLong-horizon planning, state trackingThermal effects, timing adaptation, error recovery

RoboCasa vs. Related Kitchen Benchmarks

FeatureRoboCasaHabitat 2.0BEHAVIOR-1KAI2-THOR
Kitchen focusDedicated kitchen benchmarkWhole-home (incl. kitchen)Multi-room householdMulti-room household
3D assets2,500+ kitchen-specificReal 3D scans1,000+ household300+ household
Kitchen layouts120+ configurationsScanned layouts50 scenes30 kitchens
Physics engineMuJoCo (robosuite)Habitat physicsOmniGibsonUnity physics
Manipulation focusPrimary focusNavigation + manipulationManipulation + navigationInteraction

Benchmark Profile

RoboCasa is a large-scale simulation benchmark for household robot manipulation. Built on robosuite, it provides photorealistic kitchen environments with over 150 object categories and 2,500 3D assets, evaluating robots on realistic home tasks like cooking, cleaning, and organizing.

Task Set
100 tasks across kitchen activities: opening/closing appliances, retrieving items from shelves and refrigerators, operating sinks and stoves, cleaning countertops, and multi-step cooking sequences. Tasks are parameterized across 120+ kitchen scene configurations.
Observation Space
RGB images from multiple cameras (agentview, eye-in-hand), depth maps, proprioceptive state, and natural language task descriptions.
Action Space
End-effector delta poses with gripper control on single or dual-arm robot platforms.
Evaluation Protocol
Task success rate across randomized kitchen configurations, object placements, and task variations. Multi-step tasks measured by subtask completion rate.

The Sim-to-Real Gap

RoboCasa provides photorealistic rendering and diverse kitchen layouts but simplifies material interactions (water, food, grease), thermal effects (stove heat, refrigerator cold), and deformable object physics. Real kitchens have unique layouts, non-standard appliance interfaces, and environmental conditions simulation cannot fully capture.

Real-World Data Needed

Kitchen manipulation recordings in real homes across diverse kitchen styles and appliance brands. Food preparation data with authentic material interactions. Multi-step cooking task demonstrations showing natural task sequencing and error recovery.

Complementary Claru Datasets

Egocentric Activity Dataset

Real-world kitchen activity video provides authentic cooking, cleaning, and organizing demonstrations with natural environmental variation.

Custom Kitchen Task Collection

Purpose-collected manipulation data in real kitchens with diverse layouts, appliances, and food items provides direct validation data for RoboCasa tasks.

Manipulation Trajectory Dataset

Contact-rich manipulation recordings complement RoboCasa's emphasis on household object interaction.

Bridging the Gap: Technical Analysis

RoboCasa represents the most comprehensive kitchen manipulation benchmark available. Its 2,500+ 3D assets and 120+ kitchen configurations create combinatorial diversity that challenges policy generalization. However, the diversity is still bounded by simulation — every kitchen uses the same physics engine, the same material model, and rendering from the same pipeline.

The food-related tasks highlight the biggest sim-to-real gap. Pouring, stirring, and cutting involve fluid dynamics, material deformation, and tool-material interactions that simulation approximates crudely. A policy that successfully 'stirs soup' in RoboCasa has not encountered real liquid viscosity, spoon resistance, or splashing.

The multi-step cooking sequences introduce temporal complexity. Real cooking involves natural interruptions (waiting for water to boil, checking oven temperature), adaptive timing (adjusting cooking time based on food state), and recovery from failures (saving over-seasoned food). These adaptive behaviors are absent from scripted simulation demonstrations.

Claru's ability to collect kitchen task data in real homes across 100+ cities provides the diversity RoboCasa aims to simulate — authentic kitchen layouts, real appliance interfaces, actual food preparation with true material interactions.

Key Papers

  1. [1]Nasiriany et al.. RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots.” RSS 2024, 2024. Link
  2. [2]Mandlekar et al.. RoboMimic: A Framework for Studying Robotic Manipulation Policy Learning.” CoRL 2022, 2022. Link
  3. [3]Ahn et al.. Do As I Can, Not As I Say: Grounding Language in Robotic Affordances.” CoRL 2022, 2022. Link

Frequently Asked Questions

Kitchen tasks involve material interactions simulation cannot model well — pouring liquids, cutting food, operating real appliances. Every real kitchen also has a unique layout, appliance set, and organization system that simulation parametrizes but cannot fully represent.

Scale. RoboCasa provides 2,500+ 3D assets across 120+ kitchen configurations, creating combinatorial diversity far beyond previous kitchen benchmarks. However, this simulated diversity must be validated against real-world kitchen variation to ensure policies transfer.

Claru's collectors operate in their own homes and local environments across 100+ cities. Each kitchen contributes unique layout, appliances, and objects — providing natural diversity that mirrors the combinatorial variation RoboCasa aims to simulate.

Not adequately. RoboCasa simulates pouring, stirring, and cutting with simplified physics that misses liquid viscosity, food deformation, knife-material friction, and thermal effects. A policy trained to 'stir soup' in simulation has never encountered real liquid resistance or splashing. Real food preparation data is essential for deploying kitchen manipulation policies.

RoboCasa provides 120+ configurations, which is more than previous benchmarks but far fewer than the millions of unique kitchens a deployed robot may encounter. Research suggests that policies need exposure to diverse layouts with varying cabinet positions, counter heights, and appliance locations to build spatial generalization. Real-world data from many homes provides this diversity authentically.

Get Real Kitchen Manipulation Data

Discuss authentic kitchen task data for validating RoboCasa-trained policies.