RoboVerseOrg/RoboVerse

What it solves

RoboVerse is a unified platform designed to enable scalable and generalizable robot learning. It addresses the fragmentation of robot learning environments by providing a single interface for multiple simulation frameworks, a vast collection of integrated datasets, and standardized training and evaluation code for various learning paradigms.

How it works

The project is structured into two primary layers:

• MetaSim: A simulation core that provides simulator handlers, scenario configurations, and a task registry to abstract away the differences between various simulators.
• RoboVerse-py: A content layer containing robots, scenes, tasks, and randomization tools.

It integrates a wide array of simulation frameworks (such as Isaac Lab, MuJoCo, SAPIEN, and PyBullet) and incorporates data and source code from numerous robot learning benchmarks (like RLBench, ManiSkill, and LIBERO). It also includes training and evaluation code for Reinforcement Learning (RL), Imitation Learning (IL), and Vision-Language-Action (VLA) models.

Who it’s for

RoboVerse is built for researchers and developers working on embodied AI and robot learning who need a standardized way to train, test, and evaluate their agents across different simulated environments and robot configurations.

Highlights

• Unified Simulation Interface: Supports a broad range of simulators including Isaac Lab, MuJoCo, and SAPIEN.
• Extensive Integration: Combines datasets and tasks from dozens of different robot learning projects.
• Comprehensive Learning Suite: Includes built-in support for RL, IL, and VLA training and evaluation.
• Modular Architecture: Separates the simulation core (MetaSim) from the content and learning logic.