A POMDP Treatment of Vehicle-Pedestrian Interaction: Implicit Coordination via Uncertainty-Aware Planning

Abstract

Drivers and other road users often encounter situations (e.g., arriving at an intersection simultaneously) where priority is ambiguous or unclear but must be resolved via communication to reach agreement. This poses a challenge for autonomous vehicles, for which no direct means for expressing intent and acknowledgment has yet been established. This paper contributes a minimal model to manage ambiguity and produce actions that are expressive and encode aspects of intent. Specifically, intent is treated as a latent variable, communicated implicitly through a partially observable Markov decision process (POMDP). We validate the model in a simple setting: a simulation of a prototypical crossing with a vehicle and one pedestrian at an unsignalized intersection. We further report use of our self-driving Ford Lincoln MKZ platform, through which we conducted experimental trials of the method involving real-time interaction. The experiment shows the method achieves safe and efficient navigation.