Farama-Foundation/Metaworld

What it solves

Meta-World is a benchmark for developing and evaluating reinforcement learning (RL) algorithms specifically for continuous control robotic manipulation. It provides a standardized way to test how well an AI agent can handle multiple tasks simultaneously (multi-task learning) or adapt quickly to new, unseen tasks (meta-learning).

How it works

The project provides a set of 50 different robotic manipulation tasks. It integrates with the Gymnasium API, allowing researchers to create environments where agents can be able to learn a single task, a group of tasks (MT10, MT50), or adapt to goal variations (ML1) and new tasks (ML10, ML45).

Users can run these environments in either synchronous mode (single process) for lower compute resources or asynchronous mode (isolated processes) for higher performance.

Who it’s for

It is designed for researchers and developers working on multi-task and meta-reinforcement learning for robotics.

Highlights

• Multi-Task Benchmarks: Supports learning across 1, 10, or 50 tasks simultaneously.
• Meta-Learning Benchmarks: Includes specific benchmarks for few-shot adaptation to goal variations and new tasks.
• Gymnasium Integration: Follows the standard Gymnasium API for easy environment creation and and interaction.
• Customizable: Allows users to create custom benchmarks by combining any number of the 50 available environments.