Towards Safe and Sustainable Autonomous Vehicles Using Environmentally-Friendly Criticality Metrics

This paper presents an analysis of several criticality metrics used for evaluating the safety of Autonomous Vehicles (AVs) and also proposes environmentally friendly metrics with the scope of facilitating their selection by future researchers who want to evaluate both the safety and environmental impact of AVs. Regarding this, first, we investigate whether existing criticality metrics are applicable as a reward component in Reinforcement Learning (RL), which is a popular learning framework for training autonomous systems. Second, we propose environmentally friendly metrics that take into consideration the environmental impact by measuring the CO2 emissions of traditional vehicles as well as measuring the motor power used by electric vehicles. Third, we discuss the usefulness of using criticality metrics for Artificial Intelligence (AI) training. Finally, we apply a selected number of criticality metrics as RL reward component in a simple simulated car-following scenario. More exactly, we applied them together in an RL task, with the objective of learning a policy for following a lead vehicle that suddenly stops at two different opportunities. As demonstrated by our experimental results, this work serves as an example for the research community of applying metrics both as reward components in RL and as measures of the safety and environmental impact of AVs.

Automated summary

This paper presents an analysis of criticality metrics for evaluating the safety of Autonomous Vehicles (AVs) and proposes environmentally friendly metrics. It discusses the applicability of these metrics as reward components in Reinforcement Learning (RL) and explores their usefulness for AI training.