Quantifying generalization in reinforcement learning

In three linesOpenAI releases CoinRun, a training environment measuring agents' generalization ability in reinforcement learning. The tool balances simplicity and challenge, simpler than Sonic but sufficient to test state-of-the-art algorithms.

Summary generated by Claude — human-verified