Computer Vision News - June 2018

we can remove all of the complexities of the world, fix the robots in place, and ask them to do the same thing over and over again. What robots cannot do are the simplest things that animals are capable of: walking through a park, picking up an acorn, throwing a can in a trash bin, picking up objects, etc. There are things that a two year old or even a one year old can do that our robots can’t. I think probably if the public thought about why there aren’t robots out there in the world helping us, they might realize that robots can’t do these things. Hollywood and the best YouTube clips paint a different picture. Probably their expectations are out in front of our current technological power. That said, that’s exactly what we’re working on in the field of deep learning and deep reinforcement learning for robotics. We’re trying to create robots that have those capabilities. Progress is rapid, and soon the capabilities of robots may catch up a little bit with the portrayal on YouTube and in Hollywood movies. Is there one specific capability that you, personally, would like to realize? I can say that there is one capability that I have been a part of pushing the state- of-the-art forward on that is an important capability… that is robots that can adapt to changes—either in their own bodies such as with damage, injury, or their batteries wearing down—or to changes in the environment, like if there is a natural disaster and a building has collapsed: First of all, that is an unknown, new environment that has never been seen before. Second of all, the robot could become damaged in such an environment. What you don’t want is what happens now which is the second that a robot is off of its pre- programmed game plan, it’s hopeless, lost, and ineffective. It just gets stuck. Instead, we want our robots to adapt, to either soldier on with their mission, such as find survivors, or at least, take themselves back to a repair station so we can repair them and send them back out into the world. If they are able to repair themselves, we might be able to one day send them to the galaxies… That’s true. The classic example is that the Mars rover got stuck in the sand. Despite years of the best engineers trying to get it unstuck, they could not adapt to that situation. We would rather send robots that can adapt on their own and also don’t require us to remotely program them... so that they can automatically make autonomous decisions including adapting to unforeseen situations to carry out missions. What will happen if one day robots will be able to produce new robots? I think that’s the beginning of an extremely powerful feedback cycle, where optimization can take over. . I think it will happen in software far before it happens in hardware, like robots. You might have AI that makes better AI. To some extent, that happens a little bit now, but not in any way like what I think your question is asking about. 19 Guest: Jeff Clune - UBER Guest Computer Vision News