Utilizing Reinforcement Learning for Adaptive Sensor Data Sharing Over C-V2X Communications

Vehicular networking has seen continued evolution over decades with the recently emerging paradigm of Cellular Vehicle-to-Everything (C-V2X) communications beginning to pick up momentum for adoption on today's roadways. Initial iterations of C-V2X grew from the LTE Device-to-Device framework and targeted application use cases that required the exchange of small packets of information: where a vehicle is, what it is doing, etc. Many of the next generation of use cases require the transfer of sensory data from vehicles to the edge, e.g., tele-driving, cooperative perception, computation offloading, etc. This work evaluates whether today's commercially available vehicular networking solutions can support the higher data rates required to carry this sensory data from vehicles to a Roadside Unit using a C-V2X testbed based on C-V2X Mode 4, which operates autonomously in shared spectrum. It is experimentally shown that C-V2X is capable of carrying the most common form of vehicle sensor data, images, with a frame latency of approximately 50 ms; however, these transmissions are often unreliable due to C-V2X Mode 4’s lack of adaptation capability. To mitigate this, a Reinforcement Learning (RL) problem is proposed that can adapt the transmission parameters of image frames using readily available out of band information in order to achieve a 23.3% relative improvement in effective throughput. Extensions to this RL problem are developed that allow explicit control over a desired risk tolerance, such as the probability of transmitting an image but not receiving it. Using this extension, the developed RL solution learns an adaptive transmission policy that successfully delivers 87.1% of the image frames which are transmitted (a 33.6% absolute improvement) while still maintaining the throughput advantage. Ultimately this work finds that today's commercial vehicular networking solutions are capable of supporting applications that require sensor data sharing by using RL to overcome the limitations of C-V2X Mode 4.