OpenAI is researching the learning behaviour of artificial intelligence through the game of hide & seek.

In an environment of walls, and moveable boxes, walls, and ramps, two teams of AIs competed against each other for over 500m rounds of hide-and-seek. One team (the seekers), gets rewarded for finding the hiders, while they, in turn, get rewarded for not being found.

Similarly to the real-life game the hiders get a few seconds head start. They can move all moveable objects, and lock them so that the seekers can not lock them as well.

Without any pre-training or rules, it is amazing to see the behaviours that emerged. The hiders went from randomly running around to first learning the environment, building an impenetrable fort with all moveable objects, and disabling anything that could allow the opponents to find them. The seekers got more creative, and learned how-to “surf” on top of boxes to penetrate the forts.