Navigation method trains autonomous vehicles on many risk preferences to cut collisions

Researchers led by Zhaofan Zhang describe a way to help autonomous systems navigate safely under uncertainty by exposing them to many different risk preferences during training, in a paper posted on arXiv.

Existing methods based on Implicit Quantile Networks usually train with a fixed or random risk level and then tune it by hand or with sensor signals in use. The new method treats changes in risk preference as shifts in quantile-weight distributions and, once deployed, plans against the worst-case preference distribution.

In experiments in simulated environments and on vehicle platforms, the method reduced collisions and improved task success, the authors report.