Autonomous vehicles (AV) and related self-driving technologies offer tremendous advantages to the United States across a variety of metrics, from economic development to socioeconomic benefits to research leadership to advancements in innovation for related emerging vision-based research areas. These related research areas range from (i) medical sciences, where the computer vision technologies that are instrumental to AV’s are likewise useful for cancer and tumor detection; (ii) Internet-of-Things of interconnected smart devices that aid in information collection and dissemination, where the recent vehicle-to-vehicle communication technologies [Anderson et. al.] are instrumental in pushing the state-of-the-art, and (iii) the ever-growing application areas of machine learning such as modern search engines, bank fraud detectors, and language modeling for translation, all of whose core algorithms are improved based on ongoing research in vision-based machine learning algorithms in AVs. Yet there are pitfalls to the current state of research in AV technologies as well, such as (a) possible pernicious economic impact from introduction of AV technology into the trucking industry, (b) unsafe technologies placed into consumer markets due to limited regulations, (c) opaque algorithms and networks currently in use for the majority of AV technology that limit understanding of the underlying tools and systems and (d) fragmentation of the AV standards landscape due to mismatch in inter-state regulatory frameworks.
We can consider these challenges along with the earlier advantages of AV technologies to derive a set of projections on the development of AV technologies and how these projections are likely to impact not only the current social and economic dynamics but also governments. We will tie these to current policy and legislative trends.
The guide from [NLC] presents five stages of autonomy for AVs. We briefly introduce the stages and refer to relevant research works in each of these stages.
No automation: This stage corresponds to the some of the vehicle technologies in consumer hands. With no automation, all parts of vehicle control are fully within driver control.
Function specific partial automation: Under this stage, some aspects of vehicle control have some automation to aid humans. Some examples include recent parking assist, lane assist, brake assist, and the more established cruise control for vehicles. The former three are recent innovations introduced to vehicles within the last decade that use vision and sensor technologies for partial autonomy.
Function specific full automation. In this stage, some functions of vehicle control are fully automated and require no human assistance. For example, recent full parking solutions allow a vehicle to park without any manual use. Such technology is more widespread in recent years due to improvements in conditional automation technologies.
Environment-specific full automation: This stage implements AV technologies under specific conditions for vehicles. There have been a few implementations of this type of AV in vehicles, including some trucks. Also, the Autopilot functionality in Tesla vehicles falls under this category, since it allows a vehicle to stay within its lane with automated steering, acceleration, and braking.
Full automation: This stage allows a vehicle to perform fully automated driving in a variety of conditions. While true full automation may simply not be possible due to the immense gap between human and computer perception, almost-full automation is more plausible. Under almost-full automation, vehicles can drive autonomously in almost all real-world environments with a few extreme exception. Currently, almost-full automation is primarily a research idea with limited consumer functionality due to a myriad challenges.
A range of technologies and related research should be funded to improve current research into AV technologies. Several such issues are introduced in [Canis, pp 4] and further elaborated in [USDOT, pp15]. We outline additional technologies and their potential impacts on vehicle autonomy not covered in [Canis]:
High-efficiency depth sensing: Depth sensing is an integral component of environment perception. Several approaches to depth sensing, including laser-based, radio-based, and ultrasonic-based approaches are outlined in [Anderson et. al., pp 61-66]. While degradation and complex environments are possible challenges, research has focused on more egregious issues in sensors: real-time efficiency of sensors using a combination of physical (which can be expensive) and AI-based sensors (which can be cheaper but less trustworthy) [Park et. al.].
Adversarial robustness: Recent research on the security of AI-based vision systems has indicated these vision systems are weak to adversarial attacks designed to exploit their learning methods to degrade real-time accuracy. A survey of techniques is provided in [Qayyum et. al.], where small changes in video capture can yield disastrous decisions in autonomous driving, such as misclassifying a stop sign as a different sign.
Machine learning explainability: Perhaps the most egregious issue in modern machine learning based AV technology is the lack of true explainability of ML and AI-based algorithms and models. While there is some recent research into elaborating the so-called “black box” of machine learning models in [Adadi and Berreda], current research is still limited in scope. The recent year-long NSTB investigation of a Uber self-driving related crash and death of a jaywalking pedestrian indicates explainability is still in its infancy [Gonzales].
We refer to several recent analyses on socio-economic impacts of AV technology in our analysis, and summarize their findings. The review in [McKinsey&Company] indicaets several improvements to living conditions due to AV technologies: (i) reduction in vehicle congestion due to increase in shared vehicles (a contributing factor to this is Uber’s research into AV technology for its pickup service, where autonomous taxi service would replace manned taxi service at lower cost than traditional taxi and lower overhead than personal driving due to lack of parking necessity in congested or populated cities); (ii) improvements in traffic flows due to more dynamic response times to lane transitions and vehicular movement (more recently, there have been significant proof-of-concept research into fully-automated congestion analysis and reduction on AV-equipped vehicles to reduce instances of traffic jams or stoppages); (iii) varying changes in city congestion based on AV implementation: usage of AV technology in already congested areas would see a decrease due to the Shared-and-Light concept, while fully private use of AV-technology would maintain or increase congestion with increase in zero-occupancy rides combined with consumer vehicle [ibid, pp 46]; and (iv) transition to a mobility-as-a-service economy to integrate the emerging AV-equipped fleets of industrial, commercial, and consumer vehicles; under such an economy, citizen mobility occurs through a mobile interface to various AV-equipped taxi fleets.
The impact on governments is less clear. While [McKinsey&Company] indicates regulatory pressure can yield faster results in the transition to fully electric and partially autonomous vehicles, the trends and difficulties in the underlying research makes this less clear. However, there are related effects in economic sectors that nevertheless may impact governments. A natural effect of electric autonomous vehicles is the requirement for charging stations. While the current limit of 300km could be surpassed with better battery technologies, it nevertheless requires building and maintenance of public fast chargers, changing demands on the public electric grid. This is different from gasoline pumps, since they are disconnected from the natural gas pipelines and the energy-grid at large, operating instead as independent stores of the fuel. Since this would be impractical for electric chargers, regulatory bodies are necessary to handle bad-faith actors and possible spikes in demand that can impact consumer electric supply [ibid, pp50].
We describe policy trends for the three research areas we outlined earlier, since we consider these of paramount importance to the future of safe and stable AV implementations:
[Anderson et. al.] Autonomous Vehicle Technology: A Guide for Policymakers. Rand Corporation.
[Canis] Issues in Autonomous Vehicle Deployment. Congressional Research Service.
[Gonzales] Feds Say Self-Driving Uber SUV Did Not Recognize Jaywalking Pedestrian In Fatal Crash. 11/2019 NPR. [Kumar et. al.] Law and Adversarial Machine Learning. 2018 arXiv.
[McKinsey&Company] An Integrated Perspective on the Future of Mobility.
[NLC] Autonomous Vehicles: A Policy Preparation Guide. NLC (National League of Cities) [Park et. al.] Robust sensor fused object detection using convolutional neural networks for autonomous vehicles. 2020 WCX SAE. [Qayyum et. al.] Securing Connected & Autonomous Vehicles: Challenges Posed by Adversarial Machine Learning and The Way Forward. 2019 arXiv.
[USDOT] Federal Automated Vehicles Policy: Accelerating he Next Revolution in Roadway Safety. 2016 US Department of Transportation and National Highway Traffic Safety Administration.
[Adadi and Berreda] Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI). 2018 IEEE Access.