Unmanned vehicle’s control real-time method based on neural network and selection function

摘要

The article deals with the problem of controlling an unmanned vehicle in real time. We trained a convolution neural network (CNN) to display raw pixels from a single front camera directly into commands of control. This end-to-end approach is almost optimal control based on the selection function. The system automatically remembers internal representations of necessary steps, such as detecting useful road characteristics with restrictions only based on the MPC’s controller calculating control commands as a training signal. Compared to explicit problem decomposition, such as obstacle detection, lane marking, path planing, and managment, our system optimizes all processing step simultaneously. An example of using the method on real robot is given.