Analyzing the inconsistency in driving patterns between manual and autonomous modes under complex driving scenarios with a VR-enabled simulation platform

Abstract

Purpose With the aid of naturalistic simulations, this paper aims to investigate human behavior during manual and autonomous driving modes in complex scenarios. Design/methodology/approach The simulation environment is established by integrating virtual reality interface with a micro-simulation model. In the simulation, the vehicle autonomy is developed by a framework that integrates artificial neural networks and genetic algorithms. Human-subject experiments are carried, and participants are asked to virtually sit in the developed autonomous vehicle (AV) that allows for both human driving and autopilot functions within a mixed traffic environment. Findings Not surprisingly, the inconsistency is identified between two driving modes, in which the AV’s driving maneuver causes the cognitive bias and makes participants feel unsafe. Even though only a shallow portion of the cases that the AV ended up with an accident during the testing stage, participants still frequently intervened during the AV operation. On a similar note, even though the statistical results reflect that the AV drives under perceived high-risk conditions, rarely an actual crash can happen. This suggests that the classic safety surrogate measurement, e.g. time-to-collision, may require adjustment for the mixed traffic flow. Research limitations/implications Understanding the behavior of AVs and the behavioral difference between AVs and human drivers are important, where the developed platform is only the first effort to identify the critical scenarios where the AVs might fail to react. Practical implications This paper attempts to fill the existing research gap in preparing close-to-reality tools for AV experience and further understanding human behavior during high-level autonomous driving. Social implications This work aims to systematically analyze the inconsistency in driving patterns between manual and autopilot modes in various driving scenarios (i.e. multiple scenes and various traffic conditions) to facilitate user acceptance of AV technology. Originality/value A close-to-reality tool for AV experience and AV-related behavioral study. A systematic analysis in relation to the inconsistency in driving patterns between manual and autonomous driving. A foundation for identifying the critical scenarios where the AVs might fail to react.

Insertion of Real Agents Behaviors in CARLA Autonomous Driving Simulator

“…As a novelty with respect to the aforementioned works, our approach intends to close the reality gap by inserting real behaviors into simulations through a virtual reality interface. We can find other studies that already incorporate human behavior in simulators using virtual reality, [13], [14], which also focus on the particular case of pedestrians and on the analysis of their behavior in interactions with autonomous vehicles [15], [16]. Virtual environments offer a flexible and controlled setting for conducting experiments with multiple types of vehicles and external HMI systems.…”

Artificial Intelligence and Software Modeling Approaches in Autonomous Vehicles for Safety Management: A Systematic Review

“…Even though the AV hardly ever crashes, people still often take over the driving from the AV. Also, even though the numbers show that the AV drives in risky situations, it rarely hits anything; this view means that the usual way of measuring safety, such as how much time there is before a collision, may not work well for traffic with AVs [72].…”