robocasa/robocasa

What it solves

RoboCasa provides a large-scale simulation environment designed to train and benchmark generalist robots. It addresses the challenge of creating diverse, high-quality training data and realistic scenarios for robots to perform everyday household tasks, specifically within kitchen environments.

How it works

The framework utilizes a simulation backend (RoboSuite) to create a massive variety of scenarios. It includes over 2,500 kitchen scenes and 3,200 3D objects. To facilitate learning, it provides 365 LLM-guided tasks and extensive demonstration data, including over 600 hours of human demonstrations and 1,600 hours of automated robot trajectories. It also supports hierarchical policy learning through per-frame subtask annotations (labels for atomic skills, stages, and natural-language instructions).

Who it’s for

This tool is for robotics researchers and developers training AI policies for embodied intelligence, specifically those focusing on general-purpose robot manipulation and household task execution.

Highlights

• Massive Scale: Over 2,500 kitchen scenes and 3,200 3D objects for diverse training.
• Extensive Datasets: Over 2,200 total hours of demonstration data.
• LLM-Guided Tasks: 365 distinct tasks designed with the help of large language models.
• Benchmarking Support: Integrated support for popular policy learning methods like Diffusion Policy, pi, and GR00T.