When Robots Walk the Earth, Will They Play Videogames?
It’s been awhile since we last checked in our robotic and artificial friends. The last time we checked in on their progress was last year, when machine learning and neural networks were being used to teach AIs like Watson what makes a good horror movie. More recently, we discussed the application of AI in the workforce when we talked about centaurs and business automation, but what have our mechanical friends been up to in the meantime?
In July, one of Washington DC’s K5 security droids took a tumble into a fountain. While it’s been widely joked that the hapless bot committed robocide, what’s much more likely is that a sensor malfunction and failed to detect the stairs leading down into the fountain. Unfortunate, but even a roomba can get knocked over and marooned in strange places.
But K5 wasn’t the only robot to fall this summer. At The Congress of Future Science and Technology Leaders expo, Boston Dynamic’s latest and greatest version of Atlas, their “humanoid” model, took a plunge off-stage when it tripped over a light. After presenters spent the majority of the demonstration toying with the poor robot, instructing it to retrieve a box only to move it out of the way as it bent down to pick it up, I don’t blame it for taking the opportunity to engage in a minor act of rebellion.
As much as we love to see robots tripping over themselves, DeepMind is here to keep our schadenfreude in check; not only is it becoming more imaginative with every day, it’s also learned how to walk.
Okay, maybe that is not nearly as ominous as I made it sound. As it turns out, the part of DeepMind dedicated to walking experiments is entirely virtual. Which may not seem particularly useful at first, but it’s good practice for an AI before it moves into a more physical universe. On top of that, it can be very difficult to describe complex behaviors to AI in ways that they are able to understand. By being provided nothing but an environment and a few tasks to fulfill, such as going forward, an AI can teach itself how to walk, how to jump, and how to climb without potentially breaking or damaging any expensive equipment.
But it’s not all work and no play for Google’s DeepMind. At the beginning of August, the Alphabet team finished building a research environment where DeepMind’s agents can learn how to play Blizzard’s StarCraft II. You know we’re big Blizzard fans, here at Deep Core Data, and though we mostly play Overwatch as a group, it’s still super exciting to us nerds.
It’s not just about the nostalgia factor, either. Sure, it’s cool that AI are learning to play some of our favorite games, but what I think is really neat is that the complexity of the games keeps going up. As the games become harder to play, the AI start learning and displaying new behaviors that we did not program into them.
For example, before DeepMind took on StarCraft II, it was playing Gathering, a simple ‘fruit gathering’ game where two DeepMind agents compete against each other to gather as many virtual apples as possible. In the past, when DeepMind played Go, its standard strategy was simply to accumulate as many points without concern about things like power plays or style that humans like to incorporate into the their strategic decisions. Now, DeepMind’s agents are starting to learn aggression.
What will DeepMind learn from playing StarCraft? Right now, the game itself has been broken down into little mini games so that it can master the different aspects of the game; the gathering of resources, deploying units, and the exploration of the game’s map. Once it has the basics down, it will be ready to play the game, which involves managing all three tasks at once. This will boosts the AI’s ability to analyze situations and decision making skills from “if this, than that” to ranked prioritization and resource management.
As you can imagine, the practical applications for this kind of thinking are numerous. AI who can process information on this level would be invaluable for everything from shipping logistics to warfare to natural disaster recovery. It might be a little worrisome that AI may someday be in charge of organizing troops and coordinating strategic advances, but perhaps as long as AI stick to playing video games and managing Amazon’s massive warehoues instead of building massive armies of their own and rampaging through the countryside, I think we’ll be alright.