Machine learning has taught a virtual skeleton to walk. Source: crowdAI.orgMachine learning has taught a virtual skeleton to walk. Source: crowdAI.orgNIPS 2017, the 31st Annual Conference on Neural Information Processing Systems, will include five official competitions this year. One, called “Learning to Run,” tasks participants with developing a controller to navigate a physiologically-based human model through a complex obstacle course.

Łukasz Kidziński, a postdoctoral fellow in bioengineering at Stanford University, created the contest as a way to help kids with cerebral palsy. In treating the condition, doctors often turn to muscle-relaxing surgery to improve a patient’s gait—which doesn’t always work. “The key question is how to predict how patients will walk after surgery,” said Kidziński. “That’s a big question, which is extremely difficult to approach.”

Kidziński works in Stanford's Neuromuscular Biomechanics Lab run by Scott Delp, a bioengineering and mechanical engineering professor who has spent decades studying the mechanics of the human body. Delp and his collaborators have collected data on the movements and muscle activity of hundreds of individuals as they walk and run.

That data can be used to build accurate models of how individual muscles and limbs move in response to brain signals. But brain control of complex processes like walking is not well understood.

Machine learning could help to model the brain’s movement control systems, says Delp, but for the most part its practitioners are working on self-driving cars, playing complex games or serving up more effective online advertising. Some, however, are looking for more meaningful problems to work on.

Crowdsourcing has brought 63 teams to the competition, with a total of 145 ideas submitted so far. Kidziński supplies each team with a human musculoskeletal model and a physics-based simulation environment for synthesizing external obstacles like stairs or a slippery floor, along with internal obstacles like weak or unreliable muscles. Each team is scored based on the distance its simulated human can travel through the obstacle course in a set amount of time.

In the long run, Kidziński hopes the work may also help others to design devices to assist with tasks such as carrying loads. Ideas could also be used to improve on sports techniques. But, he said, something important has already been created: a new way of solving biomechanical problems that looks to virtual crowds for solutions.