Researchers led by Sen Shen have developed a reinforcement-learning method that teaches self-driving cars to behave at intersections in a way that is easier on the human drivers around them, according to a paper on arXiv. An autonomous vehicle's choices affect the delay, braking and smoothness of nearby human-driven cars, the authors write.

The method combines planning in an imagined world model with correction from the results of actions actually carried out, so that observed rewards adjust both the value estimate and the driving policy.

In simulations at an intersection without traffic lights, it succeeded every time and reduced the variation in acceleration of following human-driven cars by 65.5% compared with PIWM and by 44.0% compared with DreamerV3. It also did better than PIWM in unfamiliar traffic, with less jerk at similar crossing times.

Source: arXiv Robotics (cs.RO)

Background

Sen Shen

Robotics researcher

Robotics researcher.