metaworld

Provides MT1, MT10, and MT50 benchmarks with one-hot task IDs appended to observations, enabling direct evaluation of policies learning multiple tasks simultaneously, as detailed in the Multi-Task Benchmarks section.

Follows the standard Gymnasium interface for environment creation and interaction, ensuring easy integration with existing RL libraries, as shown in the API examples using gym.make.

Supports both synchronous and asynchronous vectorized environments, allowing users to choose based on compute resources, illustrated in the MT10 and MT50 benchmark code snippets.

Allows researchers to build custom benchmarks by combining any selection of the 50 available tasks, offering flexibility for tailored experiments, as described in the Custom Benchmarks section.