Case-based reasoning emulation of persons for wheelchair navigation

Abstract

Objective: Testing is a key stage in system development, particularly in systems such as a wheelchair, in which the final user is typically a disabled person. These systems have stringent safety requirements, requiring major testing with many different individuals. The best would be to have the wheelchair tested by many different end users, as each disability affects driving skills in a different way. Unfortunately, from a practical point of view it is difficult to engage end users as beta testers. Hence, testing often relies on simulations. Naturally, these simulations need to be as realistic as possible to make the system robust and safe before real tests can be accomplished. This work presents a tool to automatically test wheelchairs through realistic emulation of different wheelchair users.

Methods and materials: Our approach is based on extracting meaningful data from real users driving a power wheelchair autonomously. This data is then used to train a case-based reasoning (CBR) system that captures the specifics of the driver via learning. The resulting case-base is then used to emulate the driving behavior of that specific person in more complex situations or when a new assistive algorithm needs to be tested. CBR returns user's motion commands appropriate for each specific situation to add the human component to shared control systems.

Results: The proposed system has been used to emulate several power wheelchair users presenting different disabilities. Data to create this emulation was obtained from previous wheelchair navigation experiments with 35 volunteer in-patients presenting different degrees of disability. CBR was trained with a limited number of scenarios for each volunteer. Results proved that: (i) emulated and real users returned similar paths in the same scenario (maximum and mean path deviations are equal to 23 and 10cm, respectively) and similar efficiency; (ii) we established the generality of our approach taking a new path not present in the training traces; (iii) the emulated user is more realistic - path and efficiency are less homogeneous and smooth - than potential field approaches; and (iv) the system adequately emulates in-patients - maximum and mean path deviations are equal to 19 and 8.3cm approximately and efficiencies are similar - with specific disabilities (apraxia and dementia) obtaining different behaviors during emulation for each of the in-patients, as expected.

Conclusions: The proposed system adequately emulates the driving behavior of people with different disabilities in indoor scenarios. This approach is suitable to emulate real users' driving behaviors for early testing stages of assistive navigation systems.