Meta-World+: An Improved, Standardized, RL Benchmark

2025-05-16

AI Summary

Meta-World, a widely adopted benchmark for multitask and meta-reinforcement learning, has suffered from inconsistent versioning and outdated documentation, undermining result reproducibility and cross-algorithm comparability. To address this, we systematically reconstruct and standardize the benchmark: (1) we achieve full reproducibility of all historical results through deterministic environment initialization, modular task definitions, and rigorous CI/CD-based validation; (2) we unify the API interface and task configuration paradigm, enabling fine-grained, customizable task suite composition; and (3) we release an open-source, Gym-compatible implementation built on Python, with explicit random seed control and modular architecture. The new version is publicly available as Farama-Foundation/Metaworld. This work advances benchmark design toward scientific rigor, substantially improving experimental reproducibility, cross-study comparability, and research efficiency—establishing foundational infrastructure for fair and reliable reinforcement learning evaluation.

Technology Category

• Multiagent Systems: Adversarial Agents
• Search and Optimization: Metareasoning and Metaheuristics
• Machine Learning: Reinforcement Learning

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Abstract

Meta-World is widely used for evaluating multi-task and meta-reinforcement learning agents, which are challenged to master diverse skills simultaneously. Since its introduction however, there have been numerous undocumented changes which inhibit a fair comparison of algorithms. This work strives to disambiguate these results from the literature, while also leveraging the past versions of Meta-World to provide insights into multi-task and meta-reinforcement learning benchmark design. Through this process we release a new open-source version of Meta-World (https://github.com/Farama-Foundation/Metaworld/) that has full reproducibility of past results, is more technically ergonomic, and gives users more control over the tasks that are included in a task set.

Problem

• Resolves inconsistencies in Meta-World benchmark for fair algorithm comparisons
• Improves reproducibility and usability of multi-task RL evaluation
• Enhances task customization in Meta-World+ for better benchmark design

Innovation

• Standardized RL benchmark for fair comparisons
• Open-source version with full reproducibility
• Enhanced user control over task sets