A Neuro-inspired Approach to Intelligent Collision Avoidance and Navigation
Nikolaus Salvatore, Sami Mian, Collin Abidi, Alan D. George
Abstract
This work presents both spiking and conventional neural network architectures, trained using reinforcement learning, for high-speed collision avoidance using dynamic vision sensors. Dynamic vision sensors are a novel class of event-based sensing equipment that enable extremely high sampling rates, but with unordinary, one-dimensional image data. A parallelized, event-stream simulation framework was developed to train reinforcement learning agents using the Microsoft AirSim UAV simulator. Event-stream data modeling output from a dynamic vision sensor was extrapolated from conventional frame-based camera feeds taken from the AirSim environment. Once trained, the spiking and conventional neural networks can be deployed to physical hardware with minimal modifications, as supported by previous work making use of the AirSim simulator.