OpenAI announced the release of Gym Retro on 25 May 2018, a platform for reinforcement learning research using classic video games [source]. The platform extends OpenAI Gym to support games from Sega Genesis and Atari 2600, with over 1,000 games initially available.

Gym Retro provides a standardised interface for training AI agents on commercial game ROMs. The platform includes integration files that define reward functions, done conditions, and game-specific variables for approximately 70 games at launch. OpenAI stated the platform would enable researchers to test generalisation across diverse game environments without requiring separate emulator implementations for each title.

The release includes a tool called the Integration UI, allowing users to create their own game integrations by defining memory addresses and reward structures. OpenAI published the platform under the MIT licence, making the core framework open source, though users must provide their own legally obtained game ROMs.

The announcement followed OpenAI's April 2018 Retro Contest, which challenged participants to train agents that could generalise to unseen levels of Sonic the Hedgehog games. That contest revealed limitations in current reinforcement learning approaches when agents encountered novel game states not present in training data.

OpenAI positioned Gym Retro as infrastructure for studying transfer learning and generalisation in reinforcement learning systems. The platform's support for thousands of games was intended to provide a broader testbed than previous single-game environments, though the practical research utility would depend on the quality and consistency of game integrations across the library.

The platform remains available as an open-source project, though OpenAI's subsequent research focus shifted toward language models and other domains.