Self-driving cars are decision-making systems whose complexity scales with the level of autonomy. In the highest level of autonomy, Level 5, the self-driving car should be able to provide full-time operation of all aspects of driving under different roadway and environmental conditions. The control architecture and its components can be designed and interconnected in different ways, but, at high level, the system structure of a vehicle with at least partial self- driving capabilities consists of a sensing module, a motion planning module, and a control module. The sensing and mapping module uses various sensor information, such as radar, Lidar, camera, and global positioning system (GPS), together with prior map information, to estimate the parts of the surrounding environment relevant to the driving scenario. The motion planner takes sensor input and determines a trajectory that the car should follow. Usually this task is accomplished using a model of the vehicle and the environment, which has limitations in its flexibility to adapt to uncertain sensor information. With the breakthrough of advanced data-driven machine-learning techniques, there are new opportunities in how to be able to design more robust and efficient algorithms for realizing self-driving cars. As an expansion to MERL’s set of techniques for autonomous vehicle design, we would like the UConn senior design team to setup a miniature-scale testbed, collect training datasets for a vehicle maneuvering on a road, and test or develop a deep neural network that uses the sensor data to determine a suitable vehicle trajectory.

Our team collaborated with Computer Science & Engineering 33 on this project.