Uncertainty and social considerations for mobile assistive robot navigation

An increased interest in mobile robots has been seen over the past years. The wide range of possible applications, from vacuum cleaners to assistant robots, makes such robots an interesting solution to many everyday problems. A key requirement for the mass deployment of such robots is to ensure they can safely navigate around our daily living environments. A robot colliding with or bumping into a person may be, in some contexts, unacceptable. For example, if a robot working around elderly people collides with one of them, it may cause serious injuries. This thesis explores four major components required for effective robot navigation: sensing the static environment, detection and tracking of moving people, obstacle and people avoidance with uncertainty measurement, and basic social navigation considerations. First, to guarantee adherence to basic safety constraints, sensors and algorithms required to measure the complex structure of our daily living environments are explored. Not only do the static components of the environment have to be measured, but so do any people present. A people detection and tracking algorithm, aimed for a crowded environment is proposed, thus enhancing the robot's perception capabilities. Our daily living environments present many inherent sources of uncertainty for robots, one of them arising due to the robot's inability to know people's intentions as they move. To solve this problem, a motion model that assumes unknown long-term intentions is proposed. This is used in conjunction with a novel uncertainty aware local planner to create feasible trajectories. In social situations, the presence of groups of people cannot be neglected when navigating. To avoid the robot interrupting groups of people, it first needs to be able to detect such groups. A group detector is proposed which relies on a set of gaze- and geometric-based features. Avoiding group disruption is finally incorporated into the navigation algorithm by means of taking into account the probability of disrupting a group's activities. The effectiveness of the four different components is evaluated using real world and simulated data, demonstrating the benefits for mobile robot navigation.Open Acces