AI Agent Hide and Seek – OpenAI

What is AI capable of learning? In a test done by OpenAI, hiders and seekers were placed in a virtual environment where they would play millions of games of hide and seek. The area included some walls, cubes, and a ramp. At first they would just run away, but then they learned skills like blocking doors with the cubes and locking them in place. After many failed attempts, the seekers learned how to use the ramps to get over the walls. To combat this, The hiders used the time they got at the beginning of the round to steal the ramp and lock it in place.

Why Playing Hide-and-Seek Could Lead AI to Humanlike Intelligence | by Synced | SyncedReview | Medium

Next, they were put in a more open environment where they would use longer blocks to build a fort. To prevent seekers from using ramps, they would lock them in place near the edges of the area. What happened next was the seekers would use the ramps to climb on boxes and box surf over to the hiders shelter. There was nothing stopping them from this because it wasn’t expected that they would learn this. I thought this was a nice experiment that could show just what AI is capable of learning.

OpenAI Tried to Train AI Agents to Play Hide-And-Seek but Instead They Were Shocked by What They Learned - KDnuggets