About this experiment

Lock and Key is a game of freeze tag played entirely by AI agents. Nobody scripts what the characters do. Each one starts out moving at random and learns by trial, error and reward, over millions of practice rounds.

The rules are simple. One seeker chases three runners in a 3D arena with a few obstacles to hide behind. If the seeker gets within one meter of a runner, that runner freezes. A free runner who touches a frozen teammate unfreezes them, and a rescued runner is protected for one second. A round lasts 20 seconds. If every runner is frozen, the seeker wins. If at least one runner is still free when time runs out, the runners win.

We built it to see how cooperation and competition emerge between learning agents, and to find out where the technique breaks. Every stage is measured and filmed, so the results come from recordings and test rounds rather than guesses.

Before any training, we measured a baseline. Runners that simply stand still already win 44% of rounds, because the seeker often cannot freeze all three in 20 seconds. Every later result is compared against that number.

The first lesson came early. With only the basic rewards, the runners learned nothing, because random runners almost never touch a frozen teammate. So we added a bonus for every rescue. It worked too well: by 2 million training steps the runners were getting frozen about 18 times per round, on purpose, just to be rescued again and collect the bonus. We fixed it with an equal penalty for being frozen, so that a freeze-and-rescue loop adds up to zero.

With the rewards corrected, the runners learned real teamwork against a scripted seeker. Then we let the seeker learn too. Within a few million steps it found a counter to the runners' strategy, and their win rate against it fell to 2%. The runners then relearned a different approach, running away and keeping to the walls and corners, and recovered to about a third of rounds against the strongest seeker.

This back and forth is the point of the experiment: each side's learning changes what the other side needs to learn. We are now retraining the seeker against the evasive runners, and we will publish the measured results, with charts and footage, as each stage finishes.

The experiment is not finished, and some of what we report may change as the seeker and runners keep adapting.