In the CoastRunners case published by OpenAI in 2016, the trained agent circled in an isolated lagoon repeatedly hitting three constantly respawning target points, scoring on average about 20% higher than human players, but it never completed the race.