User-Adapted Plan Recognition and Shared Control for Wheelchair Driver Assistance under Unvertainty

摘要

Elderly and physically impaired people constitute a continuously growing section of the world's population. A considerable part of this population group experiences mobility problems, even when using existing assistive devices such as powered wheelchairs and walkers. This reduces their quality of life substantially. Consequently, several research groups have decided to equip existing mobility devices with additional sensing andcomputing power in order to ease navigation and to reduce the number ofaccidents. This thesis focuses on the control and estimation algorithmsbehind such assistive robots. Powered wheelchairs are adopted as a testcase. In order to assist wheelchair drivers, the computer should know which manoeuvre the user desires to execute. However, with traditional wheelchair interfaces it is difficult to accurately specify the task the robot should execute. Consequently, the user's navigation plans are uncertain and should be recognised . Based on the estimated plans, the computer decides to which degree users should be assisted. Thisway, user and computer share control over the assistive robot. Furthermore, each wheelchair driver controls a wheelchair in a different manner, and user signals are not completely predictable and repeatable. Therefore, it is important for the control and estimationalgorithms to take uncertainty on user signals into account, as well asindividual driving characteristics.This work presents an approach to user-adapted plan recognition and user-adapted shared control that is different from previous approaches in three distinct ways. First, user plans are represented as trajectories from the current robot location to a goal location. This representation allows to estimate and model any complex driving manoeuvre. Second, uncertainty regarding user plans, user signals, and sensor signals is explicitly dealt with both during plan recognition and during shared control. A Bayesian approach is adopted to merge past and present information regarding the uncertain user plans. The probability distribution that resultsfrom merging past and present information regarding user plans may be multi-modal. Despite this multi-modal distribution, it should be continuously decided which robot actions are most appropriate. Three approaches were proposed and validated to make these assistive decisions under uncertainty, a Maximum Likelihood approach, a Maximum A Posteriori approach,and a greedy Partially Observable Markov Decision Process. Third, a keyelement in the approach is a user model that determines how probable the user's steering signals are under the assumption that the user wants to track a given mental trajectory. This user model can incorporate user-specific characteristics. The approach was validated through experimentswith a joystick interface and button interfaces, both in simulation andon a real wheelchair platform.