Collision Avoidance Assistance in UAV Teleoperation

Abstract

Unmanned aerial vehicles (UAVs) have found an increasing number of applications in recent years. However, the complexity of the task environment and operational requirements in many of these applications render fully autonomous operation rather impractical. This thesis presents a novel shared control strategy for human-in-the-loop teleoperation of UAVs. It integrates user direct teleoperation of the UAV with automatic collision avoidance assistance. In this strategy, the operator utilizes a human-machine-interface (HMI) to provide linear acceleration commands for the UAV in order to navigate it in the task environment. Simultaneously, a collision avoidance assistance algorithm modifies the operator's commands to help avoid potential collisions with obstacles in the environment. These corrective commands are obtained by formulating an optimization problem over a rolling control horizon and solving it in real-time, in the so-called model predictive control (MPC) framework. In the optimization formulation, obstacles represent restricted space that must be avoided by the UAV. The obstacle-free space manifests as a set of constraints on the states of the UAV. These obstacle-related constraints are generally non-convex in their original form, which can render the entire optimization problem non-convex. The obstacle free-space may be approximated by a convex region to avoid challenges associated with solving non-convex optimizations. Many of the contributions of this work relate to obtaining such convex approximations that are not overly conservative to unnecessarily inhibit the UAV ability to move in its task environment. Reachability analysis provides a powerful tool for identifying obstacles with the chance of collision, and creating approximate safe convex regions that are not too conservative. In this thesis, two new methods based on reachability analysis are proposed to generate such approximate convex obstacle-free space for the use in collision avoidance MPC. The first method uses backward reachability analysis to detect a subset of obstacles with chance of collision with the UAV in the MPC time horizon. Following the detection of these obstacles, the SVM algorithm is employed to construct a safe polyhedral convex region around the UAV. The second method improves on the first one by generating the approximate convex obstacle-free region based on forward reachability analysis. At each time step in the MPC horizon, separating hyperplanes between the UAV and obstacles are found that maximize the volume of the intersection of the UAV reach set and half-space produced by the hyperplanes. The convex safe region is approximated by the intersection of these half-spaces. The next contribution of the thesis focuses on accounting for uncertainty in the system in the development of the collision avoidance assistance algorithm. A novel ellipsoidal-based robust model predictive control (RMPC) for collision avoidance assistance in UAV teleoperation is presented. The main contribution here is the formation of the collision avoidance assistance in UAV under uncertainties as a convex optimization problem. Ellipsoidal approximation of reachable regions due to feasible inputs and uncertainties are derived using reachability analysis. A new convex generation method is used to approximate the obstacle-free space with a polyhedral. An inner polyhedral approximation of tightened constraints guaranteeing collision avoidance is obtained using geometrical relationship between the ellipsoidal reach set and polyhedral safe region. The final contribution of this thesis guarantees recursive feasibility of the MPC-based collision avoidance assistance formulation. A fundamental problem in fi nite-time MPC is the lack of guaranteed recursive feasibility. This could place the UAV into a state where collision with obstacles becomes unavoidable. In this thesis, new MPC-based collision avoidance assistance algorithms with guaranteed recursive feasibility for the UAV teleoperation with/without disturbances are introduced. Terminal velocity constraints are added to the MPC formulation to guarantee its recursive feasibility. A novel ellipsoidal tube-based MPC is introduced that extends collision avoidance assistance with guaranteed recursive feasibility to UAV under disturbances. In this method, an ellipsoidal approximation of the robust positively invariant (RPI) set is derived using a new RPI set approximation based on ellipsoidal techniques. Polyhedral approximation of the obstacle-free space is derived using the SVM algorithm. Ab inner polyhedral approximation of the tightened constraints is obtained using geometrical relation between ellipsoidal RPI set and polyhedral safe region. The methods proposed in this thesis are evaluated experimentally in an indoor environment using a fully-actuated UAV. The results demonstrate that the MPC-based collision assistance is highly effective in helping the operator navigate the task environment and avoid collisions with obstacles. This is in part due to the fact the methods introduced here generate a realistic convex approximation of the obstacle-free space. This allows the operator to teleoperate the UAV in the task environment without unnecessarily being hindered by a conservative approximation of this space.