Researchers aim to put 'safety-critical,' 'wireless' and 'autonomous vehicles' in same sentence

Imagine the highway of the future. Automated cars move swiftly down the road like a motorized school of fish, in constant wireless communication with each other. If one car perceives an obstacle on the road, it relays its evasive action to the other vehicles, giving them all time to respond appropriately.

If this sounds like the stuff of science fiction, that’s because it is — at least right now. Anyone who has had a cellphone conversation in a car knows about dropped calls. In a tightly packed pod of autonomous vehicles traveling at high speed, being out of touch for even for a fraction of a second could be catastrophic.

A team of faculty from UVA Engineering’s Link Lab, including Cody Fleming, assistant professor of engineering systems and environment, is developing a system that assures every autonomous vehicle in a group will always be optimally positioned to maximize safety and minimize communication loss. In effect, their goal is to turn fiction into reality.

The National Science Foundation determined the team’s approach is promising enough that it provided an $800,000 grant to develop a prototype system. Fleming’s collaborators are Kamin Whitehouse, Commonwealth Associate Professor of Computer Science, and Lu Feng, an assistant professor with joint appointments in computer science and engineering systems and environment.

“This is admittedly an ambitious undertaking,” Fleming said. “But the ability to connect autonomous vehicles wirelessly would yield tremendous gains in fuel efficiency and road capacity.”

Cody Fleming with graduate students at whiteboard in Link Lab

Cody Fleming, an assistant professor in the Department of Engineering Systems and Environment, and some of the graduate students in his Coordinated Systems Lab confer at a whiteboard in UVA Engineering’s Link Lab.

Tracking Moving Targets in a Dynamic Environment

Currently, engineers have relatively sound techniques to predict the performance of wireless networks when they are stationary. These methods fail when the network nodes are moving. For instance, vehicles passing a tractor-trailer with a large reflective surface can generate multipath interference that can cancel out the communications signal. In circumstances like this, following vehicles have no way of determining the best response to recover that signal and preserve their safety.

“There is a good reason we don’t ordinarily put the words safety-critical and wireless together,” Fleming said.

Fleming, Feng and Whitehouse are following an interlocking three-part approach. First, Whitehouse is using the standard array of sensors that autonomous vehicles use to orient themselves in the environment to generate a constantly evolving model of the vehicles’ physical surroundings.

“We now have pretty good ways to measure the world and predict what it will look like in the near future,” Whitehouse said.

Then he is using this data, combined with predicted motion paths, to estimate wireless channel characteristics along the vehicles’ trajectory.

Developing Worst-Case Scenarios

Whitehouse’s ability to produce real-time predictions of wireless channel properties forms the basis of Feng’s contribution to the project. Metrics used to evaluate wireless communication are typically based on averages. For safety critical systems, averages are inadequate because they might include only the occasional failure. The baseline needs to be the worst-case scenario.

Determining worst-case scenarios is a highly complex challenge. In essence, Feng is devising a way to translate a multitude of channel characteristics, which, because the vehicles are moving, are constantly changing, into an uninterrupted stream of real-time predictions about worst-case channel performance. “We need to design new formal verification methods that are fast and can address these dynamically changing channel characteristics,” Feng said.

Fleming put it this way: “The wireless channel prediction actually gives us different kinds of bounds on performance, for example the lower limit on the likelihood of getting a data ‘packet’ across with a delay of less than ‘X’ seconds, given a series of predicted vehicle trajectories,” Fleming said.

“Packet” is a unit of data containing information the “user” vehicle needs to proceed safely.

Providing Split-Second Control

Using Feng’s analyses, Fleming is designing the control system to position vehicles for the best — in other words, safest — channel performance.

“The control system seeks a trajectory that increases the chances of getting packets across the channel as efficiently as possible,” he said.

While this task would be relatively straightforward with a just single vehicle, having multiple vehicles makes it exponentially more difficult. Furthermore, the constant reconfiguration of individual vehicles, which itself changes the channel characteristics, adds another level of complexity to Fleming’s control challenge.

“To build an effective control system, we need not only to understand the world around us, but how that world will change based on our own manipulation of it,” Fleming said.

At this point, the team has developed a methodology that works theoretically and has begun testing it experimentally. A critical issue is reducing the computational burden of the system so that it can position vehicles optimally in real time.

“Right now, it takes a tenth of a second to run our algorithms,” Fleming said. “To be safe, the system needs to react in milliseconds.”