Robot Control: Ufuk Topcu's mathematics to improve autonomous system design
Wondering how soon you’ll be able to leave the driving to an autonomous vehicle? The statistics from companies developing driver-less cars sound pretty good at first blush. A few dozen Google-related vehicles have logged more than 1 million autonomous miles nationally in the past five years, while negotiating 600,000-plus stop lights, millions of other vehicles and 200,000 stop signs. Dig deeper into the facts, however, and the relaxing day dream starts to stall out.
In March, a Tesla Model S using the company’s first-generation auto pilot followed the original road-lane markers on a busy highway rather than reflective markers placed on the road during construction. The autonomous vehicle (with a human tester aboard) slammed the driver’s side of the car into a concrete barrier. And last February, the subsidiary of Alphabet (which also owns Google) developing autonomous cars had a major wrinkle in its 19-0 record of always being the victim in crashes rather than the cause. One of Waymo’s driverless vehicles (with two riders) assumed that a bus would yield to it after the Lexus dodged some sandbags on a Mountain View street. It struck the bus as a result.
“Whenever people point to Google’s self-driving vehicles as a great accomplishment,” says Dr. Ufuk Topcu, “I say ‘When the Wright Brothers figured out how to fly, it wasn’t the end of understanding aerospace engineering, it was the beginning.”
An ICES core faculty member who designs and evaluates autonomous systems, Topcu is particularly keen on improving computer interfaces that are used to monitor or control an autonomous device. The goal is to take into account imperfect “behavior” on both sides of a control panel. Just last month, the assistant professor of aerospace engineering and engineering mechanics was selected by the National Science Foundation for a prestigious 2017 Faculty Early Career Development (CAREER) Award. The award will fund studies of ways to reduce errors during interactions between humans and autonomous systems.
The control interface of robots and other autonomous systems traditionally are upgraded like most systems: based on the performance of past versions, and tweaks to improve features that are then beta tested. This as-you-go evaluation approach can have stark consequences when the machine in use can respond independently. For instance, a worker at a Volkswagen plant in Germany died in July 2015 after being struck in the chest while installing a robot (inside its safety cage). To address such concerns, Topcu is building a mathematical foundation for developing autonomous systems so that flaws occur less often in their ability to adapt in real-time to a task and to accommodate the vagaries of human behavior.
“It can be a person who operates a drone from thousands of miles away, or a person with a disability sitting on a kind of “smart” wheelchair,” Topcu notes. “The interactions are limited by the realities of current interfaces used to connect the two.”
Other systems whose interactive capabilities could be fine-tuned include autonomous submarines that track their foreign counterparts, distributed systems that dissipate heat in high-performance aircraft, and systems that handle chemicals at petroleum and other plants.
Topcu will focus the CAREER award on identifying and addressing limitations inherent to both sides of a computer interface – that is, not just human shortcomings, but also those of autonomous machines. For instance, people have the upper hand when it comes to inference – recognizing what elements to consider to avoid a car accident, or walking fluidly across a room after someone flips the lights off momentarily. “Robots, autonomous systems, cannot do that, fill in missing data,” Topcu says. “But if a human is managing 10 different robots, they will get tired easily.”
Topcu’s interest in improving how things work began as a child growing up in Turkey. But he readily admits not being a go-to-guy for mechanical repairs. “I theoretically know how something like a car’s engine works,” he says, “but I’m much more into the math of things.”
Scoring well on a national exam allowed him to pursue a mechanical engineering degree at the well-regarded Bogazici University. After graduating with honors, he moved to California in 2003 to pursue a master’s degree (in mechanical and aerospace engineering at UC Irvine) and a doctorate (in mechanical engineering at UC Berkeley). He began focusing at Berkeley on how to make autonomous systems more reliable and affordable. While a postdoctoral researcher at Caltech in Pasadena, Calif., he received the first recognition for his promising career: a 2012 U.S. Air Force Young Investigator Award.
ICES’ Director Tinsley Oden notes that Topcu’s intelligence and drive will likely continue to take him far. “He’s well educated, has been focused on his field for a number of years in a dedicated way, and he’s an exceptional scholar with a competitive approach that has positioned him to become a leading scientist,” Oden says.
Before joining The University of Texas at Austin faculty, Topcu held a research faculty position at the robotics-heavy University of Pennsylvania. The appeal of working at a campus with wide-ranging strengths led him and his wife, Assistant Professor Zeynep Somer-Topcu in UT’s government department, to Austin. He applies his intellect while guiding more than a dozen doctoral candidates and postdoctoral researchers in studying multiple theoretical and algorithmic aspects of designing and verifying autonomous systems.
The research on the systems’ behavioral and technical challenges incorporates knowledge from disparate fields that are often worked with in isolation. His studies include knowledge about how learning occurs, control theory, and theoretical computer science.
“We can’t pre-program everything about an autonomous vehicle, for example,” Topcu points out about the need to apply learning theory. “It’s a safety-critical vehicle that will sometimes have to respond on-the-fly.”
The work also incorporates an understanding of the behavior of dynamical systems that receive inputs and can adjust their actions based on feedback (controls theory). Part of what future autonomous devices may be guided to do, for instance, is determining at any moment how much control to assign a person based on measurements of their hand-eye coordination, alertness and other variable behaviors.
For the theoretical computer science, Topcu’s work draws on formal methods, which are mathematically based techniques to develop and verify computer systems. Combining information from different fields into a mathematical framework allows more rigorous predictions to occur of potential failures in an autonomous system’s interactions.
“He’s working at the intersection of diverse engineering disciplines, bringing them together using his analytical background to develop unique insights about how to approach the design of complex, often safety-critical, software-controlled systems,” says Richard Murray from Caltech. A professor of control and dynamical systems and bioengineering there, Murray supervised Topcu’s postdoctoral work until 2012.
Topcu’s lab ultimately uses findings to develop software for autonomous infrastructure under study. Lab members with engineering degrees often join him in adding computer science skills to their repertoire for the multifaceted projects they pursue.
As a teacher-scholar, he also incorporates undergraduates into the research wherever possible. As an example, he has dedicated a full-time laboratory assistant to help undergraduates in the university’s Women in Aerospace for Leadership and Development program as they finish a project to improve an autonomous interface. As part of the year-long project, the student volunteers have donned a brain helmet that allows them to verify how well the software adjustments they’ve made impact their ability to fly miniature drones indoors, using just their eyes to direct a drone’s movements.
The fluid nature of his studies has Topcu interacting with many overseas collaborators. He has taught short courses about distributed embedded systems in France, Italy, and Serbia. And he will host undergraduates this summer from France, Sweden, and Minnesota (the latter is participating in ICES’ Moncrief Undergraduate Summer Internship program).
Among the projects the students could participate in is investigating a robot that must navigate the visually challenging, shifting landscape of an aircraft carrier deck. Topcu also has a project to map out the pre-programmed activities that would be required of spacecraft and robots on lengthy manned missions to Mars.
“Autonomous systems can be applied in such versatile ways,” Topcu says. “There’s huge potential to grow our capabilities when it comes to developing shared-authority interfaces that will meet many societal needs.”