Deep Reinforcement Learning Based Sensor Data Management for Vehicles

Sensing the surroundings and processing the sensed data in a timely manner are critical as part of the perception process of autonomous vehicles that are getting closer to reality. This paper addresses the issues of managing the data streams generated from a variety of sensors embedded in a vehicle. To do so, we first adopt the real-time database structure that provides the clear programming abstraction over time-varying heterogeneous sensor data, and then exploit deep reinforcement learning (DRL) that can deal with highly dynamic data updates with underlying system restrictions such as limited in-vehicle network bandwidth. The experiments demonstrate that our DRL-based approach performs steadily against rapidly changing data and is able to efficiently suppress unnecessary data updates.