Many serious automobile accidents could be avoided if drivers were warned of impending crashes before they occur. Creating such warning systems by hand, however, is a difficult and time-consuming task. The goal of this project is to evolve neural networks with NEAT (NeuroEvolution of Augmenting Topologies) to warn about such crashes in real-world environments.

NEAT was used to train warning networks in a complex, dynamic simulation of both open-road driving as well as driving with other cars. Different sensor modalities were evaluated, resulting in the suprising discovery that NEAT was able to successfully generate warning networks using only raw pixel data from a simulated camera. This approach was also implemented on a real robot to determine how well this approach scales from simulation to the real world.

See vehicle warning movie page for a demo.