Meta-World — Multi-Task and Meta-Reinforcement Learning

Meta-World is a simulated multi-task and meta-reinforcement learning benchmark developed at UC Berkeley and Carnegie Mellon University. Released in 2019 under MIT license, it provides 50 distinct robotic manipulation tasks simulated with a Sawyer robot arm, covering a diverse range of tabletop manipulation operations including pushing, reaching, picking, placing, and tool use. The benchmark is designed for evaluating multi-task learning — training a single policy across all 50 tasks — and meta-learning — training policies that can rapidly adapt to new tasks. Meta-World became the standard benchmark for multi-task robot RL research, with leaderboard-tracked results enabling systematic algorithm comparison. The dataset contains over 500,000 demonstration trajectories generated by scripted policies across all task variations. Meta-World is maintained by the Farama Foundation as part of their open-source robotics benchmark suite.

Dataset specifications
Year2019
Episodes500,000
EmbodimentsSawyer (simulated)
Modalitiesrgb, proprioception
Task categoriesmanipulation, pick-and-place, inspection
Data formatjson, pkl
LicenseMIT
Accessopen — commercial use permitted
MaintainerUC Berkeley, Carnegie Mellon University
Origin countryUS

What is it?

Meta-World is a simulated multi-task and meta-reinforcement learning benchmark developed at UC Berkeley and Carnegie Mellon University. Released in 2019 under MIT license, it provides 50 distinct robotic manipulation tasks with a simulated Sawyer robot arm covering pushing, reaching, picking, placing, and tool use. Meta-World evaluates two capabilities: multi-task learning (a single policy across all 50 tasks simultaneously) and meta-learning (policies that rapidly adapt to new tasks). Maintained by the Farama Foundation.

Who is it for?

Researchers working on multi-task robot RL, meta-learning, and generalisation. The standard benchmark for algorithms that aim to learn many tasks simultaneously — a key capability for general-purpose robot deployment.

Key specifications

How it compares

The multi-task RL benchmark. LIBERO covers fewer tasks across 4 suites with better support for imitation learning and continual learning. RoboCasa covers more household-specific tasks but focuses on sim-to-real transfer. Meta-World's 50-task diversity and leaderboard make it the reference for multi-task algorithm papers.

Limitations and access notes

Simulation-only with a Sawyer arm. Tabletop manipulation only — no locomotion or mobile manipulation. MIT license permits unrestricted commercial use.

Linked professions

Frequently asked questions

What is Meta-World?

Meta-World is a benchmark of 50 simulated robotic manipulation tasks for evaluating multi-task learning (one policy on all tasks) and meta-learning (policies that quickly adapt to new tasks). Uses a simulated Sawyer robot arm, maintained by the Farama Foundation.

What is the difference between MT-10, MT-50, ML-10, and ML-45?

MT-10 and MT-50 are multi-task benchmarks where policies are trained and evaluated on 10 or 50 tasks simultaneously. ML-10 and ML-45 are meta-learning benchmarks where policies are trained on a subset and evaluated on held-out tasks, testing rapid adaptation to novel tasks.

Can Meta-World be used commercially?

Yes. Meta-World is MIT licensed, permitting unrestricted commercial use.

How do I access Meta-World?

Meta-World is available via github.com/Farama-Foundation/Metaworld and the project website meta-world.github.io. No registration is required.

Why is Meta-World maintained by Farama Foundation?

The Farama Foundation is a non-profit maintaining open-source RL environments. Meta-World was transferred to Farama to ensure long-term maintenance and compatibility with modern RL frameworks including Gymnasium.