Overview
Duckietown is a small-scale automated-driving platform where cars navigate a miniature town with pedestrians, traffic signs, and other vehicles, running computer vision and decision-making on an onboard NVIDIA Jetson board. Offloading part of that processing to a central node raises real questions about the wireless channel, given tight latency requirements and the size of the data involved.
This thesis evaluates and improves deep reinforcement learning-based driving policies under constrained communication, using both the standard simulator and a physical testbed available in the lab.
Prerequisites
Neural Networks and Reinforcement Learning