Gigaom AI Minute – July 21

Transcript

Probably the hardest difficulty roboticists face is object interaction. While robots can be physically much stronger than humans and operate in more extreme environments, at present, we are generally much more agile across a wider range of tasks. Humans have a skeletal system with two hundred bones overlaid with seven hundred or so muscles. It takes six muscles simply to move something as small as an eyeball. Replicating that kind of flexibility in a machine is hard. But even if you accomplished that, you still have the challenge of powering the creation.

To give you a sense of the difficulty roboticists are facing, consider the DARPA Robotics Challenge, which took place from 2012 to 2015. The finals were covered by Popular Science who summed it up by saying that, "the biggest and most well-funded international robotics competition in years was a failure."

The robots had to drive a car, navigate across rubble, use a doorknob, find a valve and shut it off, and so forth. They didn't have to do this all by themselves using AI. The challenge was interested in whether they could perform these physical acts, not simply whether they could perform them with no aid from humans. Additionally, the entrants knew what the robots would be asked to do. Even given these two advantages, only a few of the 24 entrants finished a course that a drunken sailor on shore leave could easily do.

It seems like an easy challenge doesn't it? What could be easier than turning a doorknob and opening the door? A lot of things it turns out. The robot has to identify the doorknob, navigate its hand to it, and squeeze it--not too hard, not too soft. The robot needs to determine the friction of the knob itself. It then needs to turn it. Humans can tell pretty easily if the knob is turning or if their hand is sliding on the knob. This is very hard for a robot. Humans can tell when to stop turning because of resistance. The robot has to be trained to actually stop at some point before it breaks the mechanism. Then, holding the knob in the open position, the robot has to push. How hard? Well, that’s extremely difficult to know ahead of time. How heavy is the door? Is it stuck? And then it has to know whether to push or pull the door.