Autonomous vehicles: An imperfect path to saving millions of lives

Autonomous vehicles (AVs) are imperfect, but they are likely to eventually become safer drivers than humans. According to the World Health Organization in 2018, 1.35 million humans died in automotive fatalities, with tens of millions more injuries and disabilities (1). Few of those deaths were the result of part failure or bad luck; the majority resulted from intoxication, texting while driving, and other distractions. Although autonomous vehicles still have a long way to develop, they already have a track record of fewer crashes than humans per million miles driven, albeit mostly under good conditions. People may disagree on the precise road conditions and safety differentials between humans and AVs, but it seems likely that eventually AVs will save millions of lives. They do not have to be perfect, in spite of the furor when one is involved in an accident. They just have to be safer, perhaps a lot safer for adoption than the currently available alternatives. AVs will transform the insurance and automotive industries, reshape transportation and delivery alternatives, and alter social behavior and the urban landscape. No amount of bad news from AV accidents over the next few years will change that outcome, independent of the timing of acceptance in different applications and jurisdictions. What remains is a lot of research, development, engineering, and testing work to continuously improve autonomous vehicles with the goal of utilizing them as soon as possible to save lives.

This special section includes a research paper titled “Neural network vehicle models for high-performance automated driving” (2). The Stanford University team trained a neural network structure using a sequence of past states and inputs motivated by a physical model. The neural network achieved better performance than the physical model when implemented in the same feedforward-feedback control architecture on an experimental vehicle. Further, when trained on a combination of data from dry roads and snow, the system was able to make appropriate predictions for the road surface on which the vehicle was traveling without the need for explicit road friction estimation. This work contributes to eventually expanding the model-based control of automated vehicles over their full operating range.

The research paper titled “AADS: Augmented autonomous driving simulation using data-driven algorithms” (3) augmented real-world pictures with a simulated traffic flow to create photorealistic simulation images and renderings. The paper, based on China/U.S. teamwork, utilized LiDAR and cameras to scan street scenes. The system generated plausible traffic flows for cars and pedestrians and composed them into the background. The composite images could then be resynthesized with different viewpoints and sensor models (camera or LiDAR). The resulting images are photorealistic and annotated. The system provides some scenarios of end-to-end training and testing of autonomous driving systems from perception to planning.

The Focus paper “Self-driving cars: A city perspective” (4) points out that, although AVs navigate cities without modifying the roadbed, they will be integrating signals from a wide variety of urban sources. These sources include traffic flows, aggregated smart phone data, car sharing, and more. In addition, AVs will drive changes in parking spaces and garages, real estate prices, commuting patterns (such as sleeper cars and buses for travel greater than 500 miles), and ride-sharing patterns. Overall, AVs will both drive and coevolve with changes in the urban landscape.

The Focus paper “Parallel testing of vehicle intelligence via virtual-real interaction” (5) reports a closed-loop parallel testing system from the longest-lasting Chinese AV competition: the Intelligent Vehicle Future Challenge of China (IVFC). Developers still lack a systematic and standardized way to test the capabilities of autonomous vehicles. The paper addressed three major programmatic objectives in testing autonomous vehicles. First, they generated and classified a standard set of driving tasks. Second, they designed a testing system that integrates simulation and field testing of real-world traffic tasks in various scenarios. Third, they built an integrated and closed-loop system to evaluate the task-specific performance of autonomous vehicles and help to improve the testing system itself.

It is interesting and useful to compare these methods with others used elsewhere. The key is not to claim superiority, but rather to encourage learning across programs over time. Additional information is available for each program via the authors. The AV industry has just begun to share data widely, with some initial caution due to competitive concerns. However, given the critical life-saving importance of AVs, we think that there is much to be gained by data sharing and continuous improvement across platforms.