Building and Evaluating an Intelligent Wheelchair

Smart wheelchairs that can detect and avoid obstacles have been developed with the goal of serving as mobility aids for persons with disabilities who find standard power wheelchairs difficult or impossible to use [58]. By improving mobility and autonomy, they have the potential to improve the health of populations ranging from the severely disabled to the growing numbers of aging people [13]. We propose to create and evaluate an Intelligent Wheelchair that represents a qualitative increase in capability, based on state-of-the-art methods for robot exploration, map-building, navigation, and directionfollowing [26, 28, 4, 39, 45]. The Intelligent Wheelchair acts under the direction of its human driver, but it is also an intelligent robot, sensing its local surroundings and maintaining a “cognitive map” of its environment. The proposed research is driven by the structure and requirements of the human-robot interface by which the human driver instructs the Intelligent Wheelchair where it should go. At the control level, the driver communicates a desired direction of motion. The wheelchair, through its sensors, maintains a model of the obstacles and other hazards in the immediate environment, and proceeds in the desired direction, deviating as needed to avoid static and dynamic obstacles. Research on the control level will include methods for multi-sensor fusion to build more accurate safety maps of the local surround. At the command level, the driver instructs the wheelchair using commands like, “Take the next left,” delivered though speech recognition, touch pad, or other clinically appropriate interface device. The Intelligent Wheelchair continues along the current path segment, identifying each significant decision point, and selects the alternative intended by the driver. Research goals at this level are to detect and describe decision points in natural indoor and outdoor settings, and to achieve human-level understanding of route instructions. At the goal level, the driver specifies the intended destination, as in “Take me to my doctor’s office.” The Intelligent Wheelchair must have a topological map of the large-scale environment, either learned from its own observations during travel experience or retrieved from a database. A topological route at this level translates naturally to a sequence of route commands, which in turn results in a sequence of hazard-avoiding control laws. Research goals include interfaces for specifying destinations, and methods for extracting topological maps from graphical maps created for humans. The higher levels of interaction require more intelligence on the part of the wheelchair, but they also require less effort for communication and supervision by the human driver. In order to maximize the human driver’s autonomy, he or she can shift freely between the different levels at any time. A preliminary version of the Intelligent Wheelchair is implemented on a robotic wheelchair in our laboratory (Figure 2). Its human-robot interface will be evaluated in three stages — alpha, beta, and gamma — with progressively more realistic environments and subject populations. Alpha evaluation tests the perception, control, mapping, and navigation capabilities of the Intelligent Wheelchair as a robot, without human subjects. Beta evaluation of the Intelligent Wheelchair takes place in virtual reality (VR) settings. This allows us to stringently evaluate hazard avoidance with human subjects, but without the risk of real hazards. Based on our previous work on VR evaluation of a mobility aid for pedestrians with low vision [59, 60], we have developed a VR version of our Intelligent Wheelchair software, allowing the driver to command the wheelchair in a physically accurate desktop-reality VR simulator. We propose to extend the fidelity of this simulation by having subjects drive the physical wheelchair in a large safe environment, while wearing a VR headset that can simulate more hazardous environments. Note that both the Intelligent Wheelchair and the human subject are receiving their perceptual input from the VR simulator.