A Bio-inspired Neural System for Energy Optimal Collision Avoidance by Unmanned Aerial Vehicles
MetadataShow full item record
In this thesis, we develop an online collision detection/avoidance approach for multi-rotor (autonomous) unmanned aerial vehicles (UAVs). One-on-one collision detection is performed by a UAV based on the observation of its own state and the state of the other UAV (via communication or sensing). The collision avoidance scheme is designed to provide a unique combination of capabilities: extremely fast execution, maximum energy conservation, and the flexibility to keep up with the planned time. Two strategic maneuvers are considered: direction change and speed change. The coherence and symmetry of the maneuver undertaken by the two UAVs is motivated by reported behavior of birds that avoid collision with each other by veering right. The online scheme is developed by performing a specialized design of (simulated) offline optimization experiments for a wide range of collision scenarios, where optimization is performed by a Mixed Discrete Particle Swarm Optimization (MDPSO). A Neural Networks classifier (to choose the best strategy when implemented online) is trained using the optimization results based an approach of constrained dominance (to assign the best approach given a particular scenario). A Neural Network prediction model is then trained using the optimization results to generate two multi-input-multi-output model that selects the best action for the two strategies in real-time based on the observed ownship/peership states. A near 100% success rate in avoiding one-one-collision is observed through evaluation with 500 test scenarios. The performance of this approach is compared with various other surrogate models and using a different optimizer.