Meta-World

Meta-World is an open-source benchmark for multi-task and meta-reinforcement learning in robotics, developed by researchers at UC Berkeley and Google Brain. It provides 50 distinct manipulation tasks built on the MuJoCo physics engine, each with a unique goal configuration spanning diverse robot skills such as reaching, pushing, pulling, opening, closing, pressing, and assembly. The benchmark is designed specifically to address the limitations of narrow task distributions in prior meta-RL research. Instead of varying only running speed or reward functions, Meta-World presents tasks with fundamentally different goals and object interactions. This enables meaningful evaluation of how well meta-learning algorithms can acquire entirely new skills, rather than just adapting to parameter variations of the same skill. Meta-World defines three core evaluation protocols: ML1 (single-task adaptation), ML10 (multi-task learning across 10 tasks), and ML45 (large-scale multi-task learning across 45 tasks). These protocols test different aspects of generalization and fast adaptation, providing a comprehensive evaluation framework that has been widely adopted in the meta-RL and multi-task RL literature. With 1,838 GitHub stars and maintained by the Farama Foundation, Meta-World is one of the most established manipulation benchmarks. It integrates with Gymnasium API, making it compatible with standard RL libraries (Stable-Baselines3, RLLib, etc.). The benchmark also supports visual observations (rendered RGB) alongside proprioceptive states, enabling research in visuomotor policy learning. Meta-World complements other manipulation benchmarks by focusing specifically on multi-task and meta-learning evaluation. While RLBench tests across 100 varied tasks with vision-based control, Meta-World provides a more controlled setting for studying how agents learn to learn across distinct skill boundaries. It remains the standard benchmark for evaluating multi-task policy generalization in robotics.