Download
Meta-World is an open-source benchmark suite of robotic manipulation environments focused on multi-task and meta reinforcement learning. It provides a large collection of continuous-control tasks, such as reaching, pushing, opening doors, and manipulating objects with a simulated robot arm. The library defines standardized benchmarks like MT1, MT10, and MT50 for multi-task learning, where a single policy is trained across different numbers of tasks. It also offers meta-learning benchmarks (ML1, ML10, ML45) that evaluate few-shot adaptation to new goals or entirely new tasks. The environments adhere to the Gymnasium API, which makes them easy to plug into existing RL pipelines, and they support both synchronous and asynchronous vectorized execution for running many environments in parallel. Installation is done via pip, with official support for Python versions 3.8 through 3.11 on Linux and macOS, and the project is licensed under MIT to encourage broad academic and industry use.
Features
- Large collection of robotic manipulation tasks for continuous control RL
- Multi-task benchmarks MT1, MT10 and MT50 for training policies across many tasks
- Meta-learning benchmarks ML1, ML10 and ML45 for few-shot adaptation evaluation
- Gymnasium compatible API with support for vectorized synchronous and asynchronous environments
- pip installable Python package supporting modern Python versions on Linux and macOS
- MIT licensed benchmark maintained by the Farama Foundation with active releases
Project Samples
