MetaWorld

Multi-task and meta RL benchmark for robotic manipulation

License

MIT

Overview

An open-source benchmark for multi-task and meta reinforcement learning focused on continuous control robotic manipulation tasks. Provides standardized protocols for meta-learning (ML1, ML10 and ML45, with 1, 10 or 45 training tasks) and multi-task learning (MT10 and MT50). Integrates with Gymnasium API and supports parallel execution.

MetaWorld is a comprehensive benchmark suite designed for evaluating multi-task reinforcement learning and meta-learning algorithms in continuous control robotic manipulation domains. The benchmark offers meta-learning protocols with 1, 10 or 45 training tasks (ML1, ML10, ML45) and multi-task protocols up to all 50 tasks (MT50), enabling researchers to systematically study how learning algorithms scale with increasing task diversity and complexity. Built with the Gymnasium API standard, MetaWorld provides familiar interfaces for RL practitioners while supporting efficient parallel execution of multiple environments. The task suite covers a diverse range of manipulation primitives and scenarios, creating opportunities to study generalization, transfer learning, and few-shot adaptation in robotic control, a critical capability for deploying adaptable robotic systems that can handle the diversity of real-world manipulation challenges.

Why it matters

MetaWorld is the most widely-used benchmark for multi-task and meta reinforcement learning in manipulation. Its 50-task suite has become the standard way to evaluate whether RL algorithms can learn and transfer across diverse manipulation skills.

Problem it solves

Multi-task RL research lacked a standard benchmark with enough task diversity to meaningfully evaluate generalization. MetaWorld provides 50 tasks with standardized protocols that reveal whether algorithms truly learn transferable manipulation skills.

Key takeaways

• 50 distinct manipulation tasks sharing a common Sawyer robot setup
• Three benchmark protocols: ML1, ML10, ML45 for different evaluation scales
• Designed specifically for multi-task and meta-learning algorithm evaluation
• Consistent observation and action spaces across all tasks
• Built on MuJoCo for fast, reliable simulation

Who it is for

• Researchers in multi-task and meta reinforcement learning
• Teams developing continual learning algorithms
• Developers evaluating task transfer and generalization
• Students studying advanced RL for manipulation

Limitations

• All tasks use the same Sawyer robot — no cross-embodiment evaluation
• Task definitions are relatively simple compared to real-world manipulation
• State-based observations dominate — vision-based evaluation is secondary
• Reward functions can sometimes be exploited without solving tasks correctly