Meta-World

A simulation benchmark of 50 robot-arm manipulation tasks, used to test multi-task and meta-reinforcement learning.

Meta-World is an open-source simulation benchmark proposed by Tianhe Yu, Chelsea Finn, Sergey Levine, and colleagues at CoRL 2019: a Sawyer robot arm performs 50 distinct manipulation tasks in MuJoCo, such as opening a drawer, pressing a button, pushing a block, or opening a window. It supports two evaluation setups: multi-task learning (MT1/MT10/MT50, learning 1, 10, or 50 tasks at once) and meta-learning (ML1/ML10/ML45, training on one set of tasks and then testing how fast the model adapts to new ones; ML45 uses 45 training tasks plus 5 held-out test tasks). The original paper found that existing algorithms already struggled to learn just 10 tasks at once. The project is now maintained by the Farama Foundation, has switched to the Gymnasium interface, and released Meta-World+ (NeurIPS 2025), which standardized version details; it remains a common benchmark for multi-task learning, meta-reinforcement learning, and imitation learning.

ExampleIn the ML45 setup, an agent is meta-trained on 45 tasks and then faces 5 tasks it has never practiced, such as a novel way of opening a door; it is scored on how quickly it can learn the new task from a small amount of interaction, measured by success rate.

Also called

MetaWorld, Meta-World+, MT10, MT50, ML10, ML45

Related

Meta Reinforcement Learning · Multi-Task Learning · MuJoCo (Multi-Joint dynamics with Contact) · Benchmark · Success Rate · Gymnasium

Sources

Meta-World: A Benchmark and Evaluation for Multi-Task and Meta Reinforcement Learning (arXiv 1910.10897)
Farama-Foundation/Metaworld (GitHub)

As of

2026-09