Applied Stochastic Models in Business and Industry has just published a special issue on Statistics of the Autonomous Vehicles, guest edited by David Banks and Feng Guo.

The foreword is reproduced below:

Technology is amazing, especially to those of us who are of a certain age. Cell phones, Google maps, streaming television, search engines, Amazon shopping, Zoom, e-books, in-flight entertainment, Overleaf, large language models and various music platforms have changed almost every aspect of our lives. The one exception is the act of driving, which still predominantly demands continuous attention and control from a human driver.

The recent advancement in autonomous vehicles indicate that the human-only driving will soon change too. Different stakeholders have different views on how that change will happen, when it will happen, what form it will take, and to what extent it can replace the role of drivers, but the consensus in the automotive industry and research community is that autonomous vehicles are coming on fast. And it seems clear that no matter what form the revolution ultimately takes, it will pose statistical issues that our profession should address.

The obvious one is safety. Risk analysis for autonomous vehicles is sensitively dependent to the way in which they are adopted. If the vehicle makes requests for a human driver to intervene when its AI encounters some circumstance outside of its confidence zone, as discussed in this special issue’s paper by Naveiro, Caballero and Rios Insua, then there are obvious questions of safety. Human beings are not reliably alert, and consequently there are ethical concerns conflated with AI training. These issues are especially acute in the mixed fleet scenario. In contrast, the fully connected autonomous vehicles described in Karimoddini et al. are likely to be safer and require less human guidance, but the legal and social path to that future is unclear.

One issue in risk analysis is to forecast the rate at which problematic driving situations arise. This is a moving target, since AIs in autonomous vehicles get smarter as their experience grows. The paper by Terres, Chen, Zhou and McLeod describe explicitly statistical technique for such inference.

The fourth paper in this special issue concerns prediction of vehicle speed trajectories. Such prediction is associated with risk, since it enables assessment of predictive control of autonomous vehicles. It is also needed for estimating fuel efficiency. Classical approaches to the problem have used Markov chains, but the paper by Behnia, Karbowski and Sokolov use generative adversarial networks to simulate realistic paths and speeds.

Of course, these papers are just the statistical starting point. The world will need to measure the economic impact of autonomous vehicles, which have the potential to both cause massive disruption of employment and huge commercial gains from inexpensive, efficient transport. The social consequences are also subject to statistical study—will autonomous vehicles promote more independence for the elderly and for children? The environmental aspects are important, as is quantifying the effects of regulation. Autonomous vehicles pose many complex and explicitly statistical problems. We challenge our profession to address them!