Madhur Behl Earns National Science Foundation Grant Fueling Research that will Lead to Safer, More Trusted Autonomous Vehicles
In May 2020, Partners for Automated Vehicle Education shared results from a poll of 1,200 Americans about attitudes around autonomous vehicle technology. Three in four believed the technology was “not ready for primetime;” almost half indicated they would never ride in a self-driving car; and a fifth do not believe that autonomous vehicles will ever be safe.
The poll outlines the deep skepticism surrounding self-driving vehicles. Methods to improve—and prove—safety will be needed for broad-based acceptance.
Madhur Behl, an assistant professor with joint appointments in the University of Virginia School of Engineering’s Department of Computer Science and Department of Engineering Systems and Environment, has earned a CAREER Award from the National Science Foundation for his pioneering research that will accelerate safety for autonomous vehicles.
Using auto racing as a platform, Behl has invented artificial intelligence methods to agilely maneuver an autonomous vehicle while pushing the limits of its steering, throttle and braking capabilities. His novel racing research is creating advanced algorithms that hold the key to safer autonomous vehicles, enabling them to avoid collisions even when they encounter unexpected challenges at high speeds while close to obstacles or other vehicles.
Behl, who serves on the Academic Advisory Council for Partners for Automated Vehicle Education, will demonstrate the power of his methods during a major milestone in the quest to advance autonomy. This October he will lead UVA’s Cavalier Autonomous Racing Team when they compete with a fully autonomous Dellara IL-15 vehicle at speeds of 180 mph on the Indianapolis Motor Speedway.
The Indy Autonomous Challenge is the first autonomous race in the world featuring life-sized racing cars.
“In many ways, the Indy Autonomous Challenge is like the DARPA Grand Challenges,” Behl said. Introduced in 2004 by the Defense Advanced Research Projects Agency to accelerate the development of autonomous vehicle technologies, the inaugural challenge tested the limits of autonomous vehicles on a 150-mile track in the Mojave Desert. It was the most extreme test of the technology for that time and led to successively grander DARPA challenges, some in simulated urban settings.
“The fact that driverless technology was gaining traction was apparent,” said Behl, who was completing undergraduate studies at the Punjab Engineering College in India during the rise of the DARPA Grand Challenges. With a life-long love of robotics, he was serving as president of his university’s robotics team. An invitation his senior year to join the robotics team at the Institute of Robotics and Intelligent Systems at ETH Zurich put Behl on the path to the Indy Autonomous Challenge.
“As an undergraduate, participating on the robotics team was my first real exposure to research. We were working on real-world problems with swarm robots, creating algorithms to control 50 different robots to move together,” Behl said. “The experience made me realize that robotics, graduate school and research was a viable and exciting career path, and I decided to pursue a doctorate.”
Behl earned his master’s and Ph.D. degrees in electrical and systems engineering from the University of Pennsylvania, where he focused on research in autonomous systems. While at Penn, he won the World Embedded Software Challenge, the Richard K. Dentel Memorial Prize for research and excellence in urban transportation and the 2011 World Embedded Software Contest.
He also co-founded the F1TENTH autonomous racing conferences in 2015. F1TENTH was modeled on traditional motorsports and hosted races that pitted autonomous vehicles, at one-tenth scale of full-size race cars, in direct competition on the track. Behl still serves as a race director, organizing F1TENTH events and speaking about the racing research platform at international conferences.
F1TENTH attracted university teams from across the globe, and more than 50 universities have adopted the racing model for their autonomous vehicle research and teaching. When Behl came to UVA Engineering in 2017, he shared the experiential lessons from the track with students in his undergraduate F1/10 Autonomous Racing course.
The course has become one of the most popular among computer science classes in UVA’s School of Engineering. Students build, drive and race autonomous cars that are one-tenth the scale of real vehicles to learn the principles of perception, planning and control that are so critical for safely operating vehicles. Each semester culminates in a head-to-head team race for a final exam. Behl shares the course curriculum with other universities.
Behl is also on the leadership council of UVA Engineering’s Link Lab, where over 40 faculty and more than 200 graduate researchers are devoted to interdisciplinary research in cyber-physical systems. One of the Link Lab’s research focus areas is autonomous systems.
Behl’s research aims to develop autonomous systems capable of safely navigating complex environments while also reacting appropriately to unexpected events. He envisions a future when autonomous vehicles are accountable for their actions and behave according to human values when faced with difficult, ethical choices while driving.
“Although there has been great progress in the last decade, and consumers can now purchase a semi-autonomous vehicle, the problem of driving is still very complex,” Behl said. “For driverless cars to be successful, they need to solve the same tasks that humans do subconsciously while driving. This is really tricky.”
Human drivers make split-second decisions in day-to-day driving, like navigating high speed merges in which nobody is yielding, slamming on the brakes when the car in front stops, or swerving when an animal runs into the road. These extreme situations exemplify the unusual conditions that pose the most difficult safety challenges for autonomous systems.
Behl points out that a majority of behavior data used to develop algorithms has been gathered through self-driving car studies that are restricted to highways, where traffic patterns are structured and predictable. Driving in urban areas is much more complex, with frequently changing variables requiring autonomous vehicles to exhibit acute perception, planning and control.
“Urban settings are a problem in designing autonomous vehicles that can deal with unexpected situations in traffic. People do not act rationally when they drive, so many things can go wrong,” Behl said. “There will always be one more thing that could not possibly have been predicted in off-highway driving.”
The key to faster adoption is finding methods to ensure autonomous cars can make safe decisions when they encounter situations that are uncommon and difficult to predict. The NSF CAREER grant supports Behl’s research around the idea that if autonomous vehicles are more agile, they can react quickly in unpredictable situations by aggressively braking or steering when a collision is imminent.
In the Link Lab, Behl is racing the one-tenth scale driverless vehicles to accelerate agile, safer designs. There are many extreme scenarios that can be created and observed in test racing. Even at the one-tenth scale, the autonomous vehicles can reach speeds of up to 16 to 20 mph.
Maneuvering the autonomous vehicles close together at high speeds forces them to use expert driving behavior and deal with the largest number of extreme challenges in short timeframes. This also generates a lot of data about expert driving behaviors. The data can be used to create algorithms—what Behl refers to as agile control algorithms—that will make autonomous vehicles capable of more expert driving behaviors.
“Autonomous racing presents unique opportunities and challenges in designing algorithms so vehicles can operate firmly on the limits of perception, planning and control. To succeed at racing, an autonomous vehicle is required to perform both precise steering and throttle maneuvers in a physically complex, uncertain environment by executing a series of high-frequency decisions,” Behl said. “This makes racing an interesting opportunity to explore the physical and algorithmic limits of autonomous driving.”
The F1/10 Autonomous Racing course offers undergraduates a window into Behl’s research methods, which have drawn increasing levels of interest from students across Grounds. His own positive experience with research as an undergraduate prompted Behl to launch the UVA Cavalier Autonomous Racing Club last year so he could offer the hands-on experience to even more students.
“As a professor I never want to turn students away from a unique learning opportunity, and one that exposes them to research. That is what led me to think that we should make autonomous motorsports a University-wide initiative for any student who has the motivation and the time,” Behl said. “It just so happened that the timing of the club launch aligned with the announcement of the Indy Autonomous Challenge, so it became our first club race.”
Behl is mentoring the team in pre-race hackathons, or virtual races using algorithms, to earn their Indy Autonomous Challenge race day placement. In June, the team will get to test their skills at the Indianapolis Motor Speedway. Then the top teams’ algorithms take center stage in October, competing head-to-head to autonomously race a Dellara IL-15 racecar across the finish line in hopes of earning a $1 million prize.
Behl believes in the power of motorsports—especially cutting-edge races like the inaugural Indy Autonomous Challenge—to transform autonomy and accelerate acceptance from a deeply skeptical public.
“Motorsport racing is where new concepts and technologies can be proven and stressed to the breaking point long before production lines are established,” Behl said. “It is worth remembering there was a time people worried about cars themselves and did not have trust in that technology. It was through racing events that car manufacturers were able to showcase the safety of their vehicles.”
He also believes future autonomy industry leaders will emerge from extreme challenges like the one taking place at the Indianapolis Motor Speedway. Behl observed the participants of the original DARPA Grand Challenges migrate into industry to head up the autonomous divisions of major car manufacturers and tech giants like Waymo, Google and Tesla.
Providing opportunities for students to define the autonomy industry—something Behl himself has done—is a major driver in advising the team.
“You can either be a bystander and just see where life takes you in the next decade, or you can partake in shaping the future,” he said. “Not only do I want to solve this real-world challenge, my goal is to mentor students in becoming the next generation of leaders in this space.”
"Madhur’s key insight is that advances in autonomous vehicle operation at racetrack speeds will allow autonomous vehicles in everyday traffic to respond more quickly and accurately. Furthermore, he is creating experiential learning opportunities that will prepare future leaders for the autonomy industry,” said Kevin Skadron, Harry Forsyth Douglas Professor of Computer Science and chair of UVA Engineering’s Department of Computer Science. “Thanks to his mentorship, UVA students will have the opportunity to develop skills for this rapidly growing market, and more generally to learn about the value of research and see first-hand how breakthroughs in the lab are critical to real-world progress.”