Delft Haptics Lab

Understanding and Reducing Conflicts between Driver and Haptic Shared Control

Abstract

Rolf Boink graduated on this Msc project, supervised by Rene van Paassen and David Abbink.

MSc thesis Abstract
Haptic driver support systems aim to guide drivers with steering using assistance torques. Many such systems generate torques using a simplified and constant lane-keeping model based on a look-ahead controller, without accounting for individual differences. Literature on haptic steering support shows beneficial effects (reduced control activity and increased time-to-lane crossings) under experimental conditions, but also report increased steering torques. However, literature has not further investigated why such increase steering torques arise.

In this study, it is hypothesized that increased interacting torques are the results of small conflicts between human and a constant haptic support system, and that these conflicts may be mitigated by adapting the parameters of the look-ahead controller to best match each individual driver behavior, essentially providing individualized guidance torques. In a fixed-base driving simulator, twenty-six subjects were asked to drive along a trajectory with constant-radius curves and straight sections, first without driver support system. Subsequently, off-line, the steering behavior of each individual driver was characterized by fitting the two parameters of the lane-keeping controller (lookahead time and gain) to best match the steering wheel angles, and these were used to generate individualized haptic guidance. In the second part of the experiment, the drivers repeated the task with the individualized guidance and with a one-size-fits-all controller based on the average controller settings of a small pilot experiment, presented in random order. After each trial, subjects were asked to state their preference.

The one-size-fits-all controller leads to higher root mean squared steering wheel torques (2.3 Nm versus 1.1 Nm for manual control) as steering into the curve, during the curve and steering out of the curve showed difference in preferred trajectory between the human and the automation. The fitting method results in individualized controller settings that significantly reduce the error between the drivers average steering behavior and the 2-parameter lane-keeping controller, but in general this does not lead to relevant improvement of the guidance as discrepancies in timing and lateral error continue to exist and is some cases even get worse. Since varying the gain and lookahead time does not lead to an overall reduction in conflict, other approaches for the driver model have to be considered.