Skip to content
MIT Department of Aeronautics and Astronautics

Aero-Astro Magazine Highlight

The following article appears in the 2007–2008 issue of Aero-Astro, the annual report/magazine of the MIT Aeronautics and Astronautics Department. © 2008 Massachusetts Institute of Technology.

Exploration, monitoring, rescue are potentials for autonomous vehicles

By Emilio Frazzoli

In little more than a year, the MIT team designed an extremely capable autonomous car, which was among only six of 89 competitors’ vehicles to successfully complete the DARPA Urban Challenge.

Emilio Frazzoli and his grad students (from left) Sameera Ponda, Joshua Bialkowski, and Byunghoon Kim in his lab with some of the vehicles they fly in their Aerial Robotics and Embedded Systems group. (William Litant photograph)

Frazzoli and students

Autonomous vehicles are a critical and rapidly developing technology both for civilian and military applications. Their use ranges from environmental monitoring and planetary exploration, to search and rescue operations, and national security. Autonomous and semi-autonomous aircraft are used extensively in military operations, and modern spacecraft exhibit ever-increasing capabilities for autonomous onboard decision-making.

The MIT Aeronautics and Astronautics Department plays an important role in the development of the state of the art in UAV control. As a doctoral student at MIT in 1997-2001, I helped develop an autonomous helicopter capable of performing acrobatic maneuvers, such as barrel rolls and split-S’s. My Ph.D. work started when my advisor, Eric Feron, and I went to a model airplane flying field with a video camera, and taped expert pilots performing impressive acrobatic feats with their remotely controlled helicopters. We decided that we would teach a computer to perform similar acrobatic routines, and to exploit their maneuverability to fly aggressively in a dangerous environment, for example, to evade threats. To do so, I devised an approach based on “motion primitives,” that is, a mathematical formalization of the intuitive concept of “maneuver.”

Such motion primitives, chosen from a library stored in the flight computer’s memory, could be combined online to construct complicated trajectories. Remarkably, this approach allowed us to rewrite the equations of motion for a complicated, high-performance vehicle in a form that resembles a kinematic system (such as a robotic arm), thus simplifying dramatically the calculations required to plan and optimize flight trajectories. Moreover, the maneuvers in the library can be designed in such a way that they can be performed reliably, thus ensuring that the vehicle will remain safe even when executing challenging maneuvers.

DARPA car

Autonomous systems have their applications for earth-bound, as well as aerial vehicles. Emilio Frazzoli worked alongside others to develop this vehicle for the 2008 Defense Advanced Research Project Agency robotic vehicle challenge.
(Jason Dorfman photograph)

 

Tackling limited interaction ability

Although the capabilities of modern autonomous vehicles are impressive, their ability to interact safely and efficiently with other vehicles (human-piloted or autonomous) is currently limited. The problem of designing vehicles that can play nice with others was at the core of last year’s DARPA Urban Challenge, a competition for robotic vehicles sponsored by the U.S. Defense Advanced Research Project Agency. MIT was one of the 89 university and industry teams from around the world that competed. Unlike previous editions of the Challenge, the racecourse was no longer primarily dirt roads, but mostly paved roads, in a road network typical of suburban areas, featuring intersections, rotaries, and parking lots. There were also winding roads and high-speed stretches.

However, the major new feature of the Urban Challenge was the introduction of traffic. Autonomous robots were expected to abide by the traffic laws and rules of the road; in principle, the autonomous vehicles were expected to display the degree of proficiency in driving skills comparable to that required to get a California driver’s license. For example, vehicles were expected to stay in the correct lane, maintain a safe speed, yield to other vehicles at intersections, pass vehicles when safe to do so, recognize blockages, execute u-turns, and park in an assigned space.

Together with Professor Jonathan How, I participated in the development of the planning and control system for the vehicle. Participation in this project was intense and rewarding. In little more than a year, the MIT team designed an extremely capable autonomous car, which was among the six vehicles that successfully completed the 60-mile race. In fact, with a fourth place finish, MIT was the first rookie in the final rankings, an achievement that made us proud. This achievement required an intense effort by all involved — undergraduate and graduate students, post-doctoral researchers, and faculty — culminating in a very emotional moment as the “start” command was sent by DARPA officials; an experience some compare to seeing a child leaving for college. At that point, we could no longer control the behavior of the car, we had to trust our prior efforts to teach the vehicle the best strategies to successfully handle the events in the race on its own.

Modern technological advances make the deployment of large groups of autonomous mobile agents with onboard computing and communication capabilities increasingly feasible and attractive. However, our understanding of such systems is still very limited. As a consequence, we are not yet close to realizing the potential offered by the ability to deploy a large number of UAVs able to perform complex missions.
Some of the limitations impeding a more widespread use of autonomous systems, both in terms of availability, and in terms of numbers of vehicles concurrently active in a shared environment, include:

  • requiring large, dedicated, and well-trained ground control crews, as well as specialized equipment
    limited ability to cope with uncertainty in a complex and dynamic environment with limited information, including on board failures, unforeseen events, and adversarial actions
  • poor understanding of the effects of scale on the complexity, performance, and cost of systems comprised of a large number of individual, (semi-)autonomous, or human-controlled units
  • limited interoperability across different vehicle systems, including human-piloted vehicles

In my current work, I concentrate on these questions for a class of problems involving a large number of mobile agents, coordinating through a wireless communication network to achieve dynamic tasks defined over a geographically extended region. For example, agents can represent mobile sensors required to collect information about a time-varying spatial field (e.g., temperature profiles, chemical concentration), or mobile relays providing wireless communication services over a region. The main objective of my work is the design of scalable, robust, and adaptive algorithms with provable performance, and with a precise characterization of the implementation complexity (e.g., in terms of computational resources, cost, or communication network bandwidth). Furthermore, I am interested in examining how the performance and complexity characteristics of the system change as its dimension grows, both in terms of the number of agents, and of the number of tasks.

Enhancing network safety, efficiency

This research requires expertise in a variety of disciplines, from systems and control theory, to optimization algorithms, distributed computation, communication networks, and operations research. With a diverse group of students from a variety of academic backgrounds, I am pursuing several projects that will ultimately enhance the safety and efficiency of autonomous vehicle networks and their ability to interact with human-piloted vehicles. Ideally, one could conceive a Turing test for autonomous vehicles, in which a human observer tries to determine which one of two operating (e.g., maneuvering in traffic) vehicles is running autonomously and which is controlled by an expert pilot. Even though much remains to be done—for example, to improve the situational awareness of autonomous vehicles, their ability to interpret the wealth of sensory information, and especially to infer the intentions of others—our recent accomplishments demonstrate that the goal of designing an autonomous vehicle able to pass such test—thus being indistinguishable from a human-controlled vehicle—is, perhaps, closer than what we could have imagined just a few months ago.


Emilio Frazzoli is an Associate Professor in the MIT Aero-Astro Department. He received the Laurea degree in Aeronautical Engineering from the University of Rome “La Sapienza” (1994), and a Ph.D. in Navigation and Control Systems from MIT (2001). Previously, he was an officer in the Italian Navy, and a spacecraft dynamics specialist at Telespazio S.p.A, in Rome. His main research interests are in planning and control of autonomous vehicles, and mobile robotic networks. He may be reached at frazzoli@mit.edu.

Massachusetts Institute of Technology, 77 Massachusetts  Avenue, 33 - 207, Cambridge, MA 02139

Contact|Site Map|Home