OpenAI has released the full version of Gym Retro, expanding its reinforcement learning benchmark platform from roughly 100 games to more than 1,000.
Gym Retro is a platform that lets researchers train and evaluate RL agents against video games using various backing emulators. The original public release covered around 70 Atari titles and 30 Sega games, a catalog that became the default proving ground for RL research. The new release pushes that count past 1,000 games across a broader range of emulators. OpenAI is also open-sourcing the internal tool it uses to add new games, giving the community a path to extend the catalog without waiting for OpenAI to do it.
Atari has been the de facto RL benchmark since DeepMind's 2013 DQN paper, but researchers have largely saturated that bar, and the narrow platform does little to test whether agents generalize across varied environments. A 10x expansion in game count, combined with multi-emulator coverage and a public intake tool, gives the field a harder and more diverse proving ground without requiring every lab to build one from scratch.
The real question is adoption. Benchmark platforms succeed when the community converges on them, and that happens through use, not through release announcements.