RoboVerse (MetaSim): Towards a Unified Platform, Dataset and Benchmark for Scalable and Generalizable Robot Learning

A robot-learning platform unifying many simulators behind one interface, with a bundled synthetic dataset and benchmark.

RoboVerse is an open-source project released in April 2025, with authors including Berkeley's Pieter Abbeel and Jitendra Malik among researchers from multiple institutions; the paper was accepted at RSS 2025, and the code is Apache 2.0 licensed. Its foundation is called MetaSim: a simulator-agnostic middleware layer that unifies backends such as Isaac Lab, Isaac Gym, MuJoCo, SAPIEN, PyBullet, Genesis, and CoppeliaSim behind the same configuration format and API, handling environment launch, asset loading, and physics stepping. This lets the same task, robots, and assets switch between different simulators, easing the problem that each simulator's formats are mutually incompatible and benchmarks are hard to compare directly. On top of this, RoboVerse pulls together tasks and data from existing benchmarks such as LIBERO, ManiSkill, RLBench, Meta-World, and robosuite, and provides a synthetic dataset, a unified evaluation protocol, and training pipelines for imitation learning, reinforcement learning, and VLA models.

ExampleWrite a grasping task's configuration once in MetaSim, then just by switching the simulation-backend parameter, run that same task with the same robots and assets in both MuJoCo and Isaac Lab, and compare how the policy performs across the two.

Also called

MetaSim

Related

Simulator · Sim-to-Sim Transfer · NVIDIA Isaac Lab · MuJoCo (Multi-Joint dynamics with Contact) · SAPIEN (SimulAted Part-based Interactive ENvironment) · Benchmark

Sources

arXiv 2504.18904 - RoboVerse
GitHub - RoboVerseOrg/RoboVerse
RoboVerse 文档站 (Chinese)

As of

2026-09