Gymnasium is described in “Gymnasium: A Standard Interface for Reinforcement Learning Environments,” posted to arXiv on July 24, 2024 by Mark Towers, Ariel Kwiatkowski, Jordan Terry, and collaborators, and accepted at the NeurIPS Datasets and Benchmarks track in 2025. It is the actively maintained successor to OpenAI Gym, now stewarded by the Farama Foundation after OpenAI stopped maintaining the original library.
OpenAI Gym had established a simple, near-universal convention for how a reinforcement learning agent talks to an environment: reset to a starting state, take an action, and receive back a new observation, a reward, and a signal of whether the episode has ended. That interface became the lingua franca of RL code. Gymnasium preserves and modernizes it, fixing long-standing inconsistencies and providing a clean, stable API that allows environments and training algorithms to interoperate freely.
Because so much RL software and so many published baselines are written against this interface, Gymnasium is effectively shared plumbing for the field. For a general reader, it shows how the long-term value of a piece of research software can come less from any single algorithm than from a widely adopted standard that everything else plugs into.