OpenAI trained AI agents to play hide-and-seek in simulation. Over millions of rounds, hiders learned to block doors with boxes and build shelters, while seekers discovered they could use ramps as launching pads. The agents kept inventing new strategies to counter each other, eventually exploiting physics in ways the designers didn't anticipate.
OpenAI built a hide and seek game for their AI agents to play. While we look at the exact rules here, I will note that the goal of the project was to pit two AI teams against each other and hopefully see some interesting emergent behaviors. And boy, did they do some crazy stuff. The coolest part is that the two teams compete against each other and whenever one team discovers a new strategy, the other one has to adapt. Kind of like an arms race situation. And it also resembles generative adversarial networks a little. And the results are magnificent, amusing, weird. You'll see in a moment. These agents learn from previous experiences and to the surprise of no one, for the first few million rounds, we start out with pandemonium. Everyone just running around aimlessly without proper strategy and semi random movements, the Seekers are favored and hence win the majority of the games. Nothing to see here. Then, over time, the Hiders learned to lock out the Seekers by blocking the doors off with these boxes and started winning consistently. I think the coolest part about this is that the map was deliberately designed by the OpenAI scientists in a way that the Hiders can only succeed through collaboration. They cannot win alone and hence they are forced to learn to work together, which they did quite well. But then something happened. Did you notice this pointy doorstop shaped object? Are you thinking what I am thinking? Well, probably. And not only that, but about 10 million rounds later, the AI also discovered that it can be pushed near a wall and be used as a ramp. And ta da gotem. The Seeker started winning more again. So the ball is now back on the court of the Hiders. Can you defend this? If so, how? Well, these resourceful little critters learned that since there is a little time at the start of the game when the Seekers are frozen, apparently during this time they cannot see them. So why not just sneak out, steal the ramp and lock it away from them? Absolutely incredible. Look at those happy eyes as they are carrying that ramp. And you think it all ends here? No, no, no, not even close. It gets weirder. Much weirder. When playing a different map, a Seeker has noticed that it can use a ramp to climb on the top of a box. And this happens. Do you think Couchsurfing is cool? Give me a break. This is box surfing. And the scientists were quite surprised by this move, as this was one of the first cases where the Seeker AI seems to have broken the game. What happens here is that the physics system is coded in a way that they are able to move around by exerting force on themselves. But There is no additional check whether they are on the floor or not, because who in their right mind would think about that? As a result, something that shouldn't ever happen does happen here and we are still not done yet. This paper just keeps on giving. A few hundred million rounds later, the hiders learned to separate all the rams from the boxes. Dear fellow scholars, this is proper box surfing defense. Then lock down the remaining tools and build a shelter. Note how well rehearsed and executed this strategy is. There is not a second of time left until the seekers take off. I also love this cheeky move where they set up the shelter right next to the seekers and I almost feel like they are saying yeah, see this here? There is not a single thing you can do about it. In a few isolated cases, other interesting behaviors also emerge. For instance, the Hiders learn to exploit the physics system and just chuck the ramp away. After that the seekers what? What just happened? But don't despair and at this point I would also recommend that you hold on to your papers because there was also a crazy case where a seeker also learned to abuse a similar physics issue and launch itself exactly onto the top of the hiders. Man, what a paper. This system can be extended and modded for many other tasks too, so expect to see more of these fun experiments in the future. We get to do this for a living and we are even being paid for this. I can't believe it. In this series my mission is to showcase beautiful works that light a fire in people and this is no doubt one of those works. Great idea, interesting, unexpected results, crisp presentation. Bravo. OpenAI. Love it. So did you enjoy this? What do you think? Make sure to leave a comment below. Also, if you look at the paper, it contains comparisons to an earlier work we covered about intrinsic motivation, shows how to implement circular convolutions for the agents to detect their environment around them, and more. This episode has been supported by Weights and Biases. Weights and Biases provides tools to track your experiments in your deep learning projects. It can save you a ton of time and money in these projects and is being used by OpenAI, Toyota Research, Stanford and Berkeley. In this blog post they show you how to use their system to find clues and steer your research into more promising areas. Make sure to Visit them through wandb.com papers wandb.com papers or just click the link in the video description and click and sign up for a free demo today. Our thanks to WeightsAndBiases for helping us make better videos for you. Thanks for watching and for your generous support and I'll see you next time.
