Gym4ReaL: Towards Real-World Reference Environments for Reinforcement Learning

ICLR 2026 Conference Submission17058 Authors

Keywords: Reinforcement learning, Real-world, Reference Environments, Datasets, Benchmark

TL;DR: Gym4ReaL: a suite of realistic reference environments designed to support the development and evaluation of reinforcement learning algorithms in real-world scenarios.

Abstract: In recent years, \emph{Reinforcement Learning} (RL) has achieved remarkable progress, reaching superhuman performance across a variety of simulated environments, largely driven by the adoption of standardized training suites, such as Gymnasium and MuJoCo. However, this success has not been translated directly to real-world domains, which present inherent challenges that remain underexplored in existing reference environments. This gap highlights the need for training suites that more closely reflect real-world conditions and facilitate the practical deployment of RL solutions. Towards this goal, in this paper, we introduce \texttt{Gym4ReaL}, an open-source suite of realistic environments developed starting from collaborations with industry partners and domain experts. The suite offers a diverse collection of tasks, simulators, and datasets that expose algorithms to real-world complexities and support the investigation of different methodological approaches. Through benchmark experiments, we demonstrate that standard RL algorithms remain competitive against expert-guided rule-based baselines in these settings, motivating the development of new methods capable of fully harnessing RL’s potential for real-world applications.

Primary Area: datasets and benchmarks

Submission Number: 17058