Autonomous vehicle navigation based in a hybrid methodology: model based and machine learning based

For decades researchers have attempted to make the car drives autonomously. One of the challenges of developing this kind of system is the environment detection and understanding where the car is supposed to drive, and control the vehicle on a safe way, based upon on the perception of the environment, provided by on-boarded sensors (cameras, LIDAR). The existing navigation methods applied to solve this problem can be organized in two categories: those based on “geometrical or physical models” and those based upon on “machine learning models”. Machine Learning based models can have good performance when trained appropriately, however, due to its “black-box” characteristics (lack of physical meaning), it can return non accurate output or even wrong values in unseen situations, leading the vehicle to collision. In the other hand, Geometric or physical approaches are designed under certain approximately modeling assumptions, based on physical/geometrical parameters that are uncertain or can not be identified, and becomes a painstaking work as it gets more complex. This paper presents a hybrid autonomous navigation methodology, which takes advantage of the learning capability of Machine Learning (ML) models, and uses the safeness of the Dynamic Window Approach geometric method. Using a single camera and a 2D LIDAR sensor, the proposed method actuates as a high level controller, finding optimal vehicle velocities to be applied by a low level controller. The system algorithm is validated on CARLA Simulator environment, where a vehicle coupled by this system proved to be capable of achieving the following tasks: Lane keeping and obstacle avoidance.