RoboVerse
Unified meta-platform wrapping 11 robotics simulators under a single API.
License
Apache-2.0
Overview
RoboVerse is a unified platform, dataset, and benchmark for scalable and generalizable robot learning. It provides MetaSim — a simulator-agnostic abstraction layer wrapping 11 simulator backends (including MuJoCo, Isaac Lab, PyBullet, SAPIEN, Genesis, CoppeliaSim) behind a single API. Includes pre-configured robots, 276 task categories, 5,500+ object assets, and 510,000+ trajectories from 18 existing benchmarks (RLBench, ManiSkill, LIBERO, Meta-World, robosuite). Integrates training pipelines for both imitation learning (ACT, Diffusion Policy, VLA) and reinforcement learning (PPO, TD3, SAC).
RoboVerse is a comprehensive open-source framework that unifies robot learning across simulation platforms, datasets, and benchmarks. Developed by a team of 37+ researchers from UC Berkeley, Peking University, USC, Stanford, CMU, and other leading institutions — with senior leadership from Pieter Abbeel and Jitendra Malik — RoboVerse tackles the fundamental fragmentation problem in robotics simulation. At its core is MetaSim, a three-layer infrastructure that abstracts diverse simulation environments into a universal interface through a simulator-agnostic configuration system (MetaConfig) and standardized API. MetaSim supports 11 simulator backends including Isaac Lab, Isaac Gym, MuJoCo, SAPIEN, PyBullet, Genesis, cuRobo, PyRep/CoppeliaSim, and Blender. The framework is organized into three pillars: MetaSim (the core simulation abstraction layer), RoboVerse Pack (pre-configured robots, tasks, and scene assets with 5,500+ assets and 276 task categories), and RoboVerse Learn (integrated learning algorithms for imitation learning and reinforcement learning). RoboVerse enables cross-simulator validation, cross-embodiment transfer, and hybrid simulation, while providing over 510,000 manipulation trajectories and 50+ million high-quality state transitions for training. Accepted at RSS 2025 and licensed under Apache 2.0.
Why it matters
RoboVerse addresses the single biggest bottleneck in robot learning research: fragmentation across simulators, datasets, and evaluation protocols. Before RoboVerse, researchers had to rewrite their code from scratch when switching between simulators like MuJoCo and Isaac Sim, making reproducibility nearly impossible. No other project provides a truly simulator-agnostic abstraction layer that lets you train a policy in one simulator and seamlessly validate it across others. By unifying 11 simulator backends, 18 benchmark datasets, and standardized evaluation protocols under a single API, RoboVerse makes it possible for the first time to conduct apples-to-apples comparisons of robot learning algorithms across different simulation backends and robot embodiments.
Problem it solves
Robotics simulation is deeply fragmented: each simulator (MuJoCo, Isaac Sim, SAPIEN, PyBullet, etc.) has its own API, configuration format, and data conventions. Researchers must rewrite entire codebases when switching simulators, making cross-platform validation impractical and reproducibility nearly impossible. Benchmark datasets are similarly siloed with incompatible formats. RoboVerse solves this by providing a universal abstraction layer that standardizes the interface to all major simulators, unifies dataset formats, and establishes consistent benchmarks — enabling researchers to write code once and run it anywhere, with less than 5% computational overhead compared to native simulator performance.
Key takeaways
- Unifies 11 simulation backends (Isaac Lab, MuJoCo, SAPIEN, PyBullet, Genesis, and more) under a single API with less than 5% overhead versus native performance
- Provides 276 task categories, 5,500+ assets, 510,000+ manipulation trajectories, and 50+ million state transitions for robot learning
- Integrates 18 established benchmark datasets (RLBench, ManiSkill, LIBERO, Meta-World, robosuite, GraspNet) with unified formatting
- Supports cross-simulator validation, cross-embodiment transfer, and hybrid simulation combining different physics engines
- Includes RoboVerse Learn with built-in algorithms for imitation learning (ACT, Diffusion Policy, VLA) and reinforcement learning (PPO, TD3, SAC)
- Adding a new simulator backend requires implementing only approximately 300 lines of code via the MetaSimInterface
Who it is for
- Robotics researchers working on imitation learning, reinforcement learning, and sim-to-real transfer
- Machine learning engineers building robot foundation models who need large-scale diverse training data
- Academic labs that need reproducible benchmarks for comparing algorithms across simulators
- Robotics startups looking to reduce simulation setup time and data collection costs
- Students and educators learning robot manipulation and policy training across multiple platforms
Limitations
- Requires NVIDIA GPU with CUDA support; RTX 3090 or better recommended for full functionality
- Isaac Sim backend requires specific NVIDIA drivers (525+) and RT Core GPUs with 8GB+ VRAM
- Windows support not yet available — currently limited to Ubuntu 20.04+ and macOS 12+
- Large storage requirements (100GB+ for full dataset and simulator installations)
- Cross-simulator physics consistency depends on individual simulator implementations, which can produce subtly different behaviors
- Still in alpha release (v1.0.0-alpha) — API may change