OpenAI announced a contest on 5 April 2018 challenging participants to develop reinforcement learning agents capable of generalising across unseen video game levels [source]. The Retro Contest focused on custom levels from Sonic The Hedgehog, requiring algorithms to perform on stages they had never encountered during training.

The contest highlighted a persistent limitation in contemporary reinforcement learning systems: models trained on specific environments frequently fail when confronted with novel scenarios sharing similar structure. Participants received access to OpenAI's Gym Retro platform, which provided an interface to classic games, but were restricted from training on the test levels themselves.

OpenAI structured the competition in two rounds. The first round ran through 5 June 2018, evaluating agents on a hidden test set of Sonic levels. A second round followed immediately, introducing levels from Sonic The Hedgehog 2 to assess cross-game generalisation. Prize money totalled $10,000.

The contest format exposed a fundamental challenge in AI capability claims. While reinforcement learning agents had demonstrated superhuman performance on fixed game environments, their inability to transfer learned strategies to modified versions of the same game revealed brittleness in the underlying systems. An agent mastering one Sonic level provided no guarantee of competence on a structurally similar but visually distinct level.

OpenAI positioned the contest as a research benchmark rather than a solved problem. The announcement acknowledged that existing algorithms struggled with generalisation, framing the competition as an open challenge to the research community. The contest structure itself constituted an implicit acknowledgement that state-of-the-art reinforcement learning systems in 2018 lacked robust transfer capabilities across even minor environmental variations.