For most animals, walking is instinctual and they're on their legs just minutes out of the womb. For humans, and for robots, it's a trickier proposition that takes a little bit of learning. But with deep learning to help out, a software robot can brave all kinds of obstacles after just a little practice, and someday real-life robots might use the same tactics.

The DeepLoco project—by Xue Bin Peng, Glen Berseth and Michiel van de Panne of University of British Columbia and KangKang Yin of National University of Singapore—is series of experiments in deep-learning locomotion presented Siggraph 2017, a conference for advanced computer animation.

In its simplest terms, the DeepLoco project has two parts. A lower-level chunk of code controls how the model walks in the most fundamental sense—how to put one foot in front of the other to head in the intended direction while meeting some basic criteria in regards to balance and style. A second, higher-level controller looks at the broader world the two-legged software bot is placed in, and uses that information to feed instructions to the one-foot-in-front-of-the-other code. Together, both trained by deep learning, they come up with results like this:

This is far from the first project designed to let simulated creatures teach themselves how to walk. Michiel van de Panne previously took part in a study where simulations allowed computers puzzle out how to animate complex and unusually-shaped creatures with some uncanny results. A vaguely kangaroo-shaped creature, for instance, opted to walk at lower speeds, but hop at higher ones:

Aside from the obvious appeal here—watching silly software robots run into things and fall on their faces—this sort of technology could prove extremely valuable in the animation of actual, physical robots in the real world. Simulations in software could surface strange-but-efficient means of locomotion that a robot's human designers might never have thought of, or help to design robots intended for very specific purposes from the ground up. It's a clever bit of engineering that will give us all something to regret while the robots chase us down.

Source: Popular Mechanics, UBC via Motherboard, Michiel van de Panne