Towards Evolved Time to Contact Neurocontrollers for Quadcopters

摘要

Bio-inspired controllers based on visual odometry - or time to contact - have been previously shown to allow vehicles to navigate in a way that simultaneously specifies both the spatial waypoint and temporal arrival time at the waypoint, based on a single variable, tau (τ). In this study, we present an initial investigation into the evolution of neural networks as bio-inspired tau-controllers that achieve successful mappings between τ and desired control outputs. As this mapping is highly nonlinear and difficult to hand-design, an evolutionary algorithm is used to progressively optimise a population of neural networks based on quality of generated behaviour. The proposed system is implemented on Hardware-in-the-loop setup and demonstrated for the autonomous landing of a quadcopter. Preliminary results indicate that suitable controllers can be successfully evolved.