Uncertainty-Aware Online Merge Planning with Learned Driver Behavior

TL;DR

This paper introduces an uncertainty-aware planning framework for autonomous merging that models and estimates driver cooperation levels in real-time, improving safety and efficiency in dynamic traffic scenarios.

Contribution

It presents a novel approach combining particle filtering and POMDPs to explicitly model and incorporate driver behavior uncertainty into merging planning.

Findings

Outperforms behavior-agnostic methods in simulations

Effectively estimates driver cooperation levels online

Enhances safety and efficiency in merging maneuvers

Abstract

Safe and reliable autonomy solutions are a critical component of next-generation intelligent transportation systems. Autonomous vehicles in such systems must reason about complex and dynamic driving scenes in real time and anticipate the behavior of nearby drivers. Human driving behavior is highly nuanced and specific to individual traffic participants. For example, drivers might display cooperative or non-cooperative behaviors in the presence of merging vehicles. These behaviors must be estimated and incorporated in the planning process for safe and efficient driving. In this work, we present a framework for estimating the cooperation level of drivers on a freeway and plan merging maneuvers with the drivers' latent behaviors explicitly modeled. The latent parameter estimation problem is solved using a particle filter to approximate the probability distribution over the cooperation level. A partially observable Markov decision process (POMDP) that includes the latent state estimate is solved online to extract a policy for a merging vehicle. We evaluate our method in a high-fidelity automotive simulator against methods that are agnostic to latent states or rely on $\textit{a priori}$ assumptions about actor behavior.