Constraint-based navigation for safe, shared control of ground vehicles

Abstract

Human error in machine operation is common and costly. This thesis introduces, develops, and experimentally demonstrates a new paradigm for shared-adaptive control of human-machine systems that mitigates the effects of human error without removing humans from the control loop. Motivated by observed human proclivity toward navigation in fields of safe travel rather than along specific trajectories, the planning and control framework developed in this thesis is rooted in the design and enforcement of constraints rather than the more traditional use of reference paths. Two constraint-planning methods are introduced. The first uses a constrained Delaunay triangulation of the environment to identify, cumulatively evaluate, and succinctly circumscribe the paths belonging to a particular homotopy with a set of semi autonomously enforceable constraints on the vehicle's position. The second identifies a desired homotopy by planning - and then laterally expanding - the optimal path that traverses it. Simulated results show both of these constraint-planning methods capable of improving the performance of one or multiple agents traversing an environment with obstacles. A method for predicting the threat posed to the vehicle given the current driver action, present state of the environment, and modeled vehicle dynamics is also presented. This threat assessment method, and the shared control approach it facilitates, are shown in simulation to prevent constraint violation or vehicular loss of control with minimal control intervention. Visual and haptic driver feedback mechanisms facilitated by this constraint-based control and threat-based intervention are also introduced. Finally, a large-scale, repeated measures study is presented to evaluate this control framework's effect on the performance, confidence, and cognitive workload of 20 drivers teleoperating an unmanned ground vehicle through an outdoor obstacle course. In 1,200 trials, the constraint-based framework developed in this thesis is shown to increase vehicle velocity by 26% while reducing the occurrence of collisions by 78%, improving driver reaction time to a secondary task by 8.7%, and increasing overall user confidence and sense of control by 44% and 12%, respectively. These performance improvements were realized with the autonomous controller usurping less than 43% of available vehicle control authority, on average.