AI research is not always the stuff of cool science fiction stories that we have been reading since we were kids. Getting to a level of perfection often involves a lot of stumbles and repeated trials on the way, and sometimes these attempts can be outright funny, even if they indicate a change or an advancement that is eventually going to transform the world as we know it. Today, the title for that weird and funny stumble belongs to Google‘s AI subsidiary Deepmind’s latest research paper “Emergence of Locomotion Behaviours in Rich Environments” or rather, more specifically, the demonstration of its results.
The paper deals with the area of AI locomotion – teaching robots how to move around obstacles in the real world more naturally, and possibly, entirely on their own. It is a fundamental aspect of reaching a level of perfection in AI research that leads to real world robots that can navigate your house, your workplace and even the streets unaided, fulfilling the objective they were designed for without human intervention. The technique used here is reinforcement learning, which is a machine learning methodology inspired by behavioural psychology. Software agents are directed to take any actions possible to maximise a potential reward that has been already specified. And it certainly works – just take a look.
Everything that you saw those stick figures doing has been achieved by themselves. The only human input in the process has been a bunch of sensors that help the agent figure out its orientation and then defining a reward as moving forward. The rest of the activities that you see – jumping, leaning, bending, even limboing in certain cases – has been devised and learnt by the AI on its own through trial and error.
The unique thing about this research is that traditionally, reinforced learning has been associated with the fragility of the knowledge gained. For instance, the agent would figure out how to climb a certain set of stairs in your home but would be entirely perplexed at the sight of an escalator. Of course, the idea of trial and error would possibly be too costly in the real world. But what this research is proving is that AI is capable of immediately correcting its ways in a complex environment and figure out robust locomotion methods and actions to interact with it and get a reward.
In the not so distant future, we could be looking at the learning in this area being implemented in multiple types of real world robots like rescue bots, deep sea diving bots and more. Who knows, someday, our very own Terminator might use this to learn how to climb all the walls we build to keep it out? Was that too pessimistic? Ok, we’ll leave it at that.