Learning Machines

Over at EDN News:

Boston Dynamics released a new demonstration video of its larger quadruped robot operating in field conditions. I am very excited by what I see in the video as it demonstrates how far the technology has progressed over the past few decades of research. In addition to the lower quality but faster-loading demonstration video, the company’s website offers a high-quality video. There are other high-quality videos of other autonomous robots at the website, including similar robots with four and six legs.

Of particular interest in the BigDog video is that much of it takes place in field conditions with slopes and real-world ground conditions, such as going up and down hills covered with leaves or even snow. There are a few slow-motion replays in the video that highlight particularly interesting behavior by the system. The first slow-motion sequence involves a person giving the robot a strong lateral shove with their foot; we see how the robot recovers from it. The next slow-motion sequence is even more impressive, as the robot walks onto an ice patch and recovers; of special note to me was how the robot used its “knees” to recover from sliding on the ice. The video finishes up with some lab demonstrations of this 340 pound (when fully loaded) robot traversing a cider block mound, a rock bed, and leaping over a section of the floor.

Behaviors like these are too complicated to program explicitly; they rely on the system learning how to interact with the world and compensate for the variability in the world.