ML, MAP and Greedy POMDP Shared Control: Qualitative Comparison of Wheelchair Navigation Assistance for Switch Interfaces

摘要

This paper describes and compares three approaches for providing navigation assistance to powered wheelchair users: a Maximum Likelihood (ML) approach, a Maximum A Posteriori (MAP) approach, and a greedy Partially Observable Markov Decision Process (POMDP) approach. The approaches are evaluated qualitatively by controlling a wheelchair using a switch interface, both in simulation and on a real setup. The results show that for this experimental setup (1) all three approaches allow the driver to reach any of the specified goal positions with greater accuracy and faster than without assistance, (2) ML produces paths that are jagged, because its decisions are based on the latest user signals only, (3) MAP decisions are much less impulsive than ML, except at the start, (4) greedy POMDP is even more cautious in taking actions prematurely because it considers the probability of all driver plans when evaluating the effect of an action.