A team from MIT has advanced to the next stage of the prestigious DARPA Robotics Challenge (DRC), a competition sponsored by the Department of Defense to promote innovation in robotics technology for disaster-response operations. The MIT team beat out more than one hundred other teams in the first stage of the competition — the Virtual Robotics Challenge (VRC) — to gain one of seven prized spots in the next stage of the competition, slated for December.
“We are thrilled to have made it this far,” says MIT team leader Seth Teller, a professor in the MIT Department of Electrical Engineering and Computer Science (EECS) and a principal investigator at the MIT Computer Science and Artificial Intelligence Lab (CSAIL). “Our team hit the ground running, building on the infrastructure that we developed for the DARPA Urban Driving Challenge back in 2007, which involved developing a self-driving car that could traverse city traffic.”
The MIT team is competing in the software development portion of the competition, known as “Track B.” The goal for participating teams is to develop software that can operate a DARPA-supplied humanoid robot across a low-bandwidth network — the only type of network that might be available to first responders in a natural disaster or other type of emergency. In the VRC, each team demonstrated the capability of its system, including its software and human operator interface, by subjecting it to a simulation environment mandated by DARPA.
“The disaster response scenario is technically very challenging,” says Russ Tedrake, a professor at EECS and a principal investigator at CSAIL. “It requires the robot and human operator to simultaneously perceive and gain an understanding for a complex, new environment, and then use that information to perform difficult manipulation tasks and traverse complex terrains.”
To address the technical challenges involved in developing disaster response control software, the MIT team is drawing on its expertise in robotic planning and control, as well as machine perception and human-robot interaction. “Our approach combines the human operator and robot body into a new kind of control system,” Teller says. The robot gathers raw sensor data and shares it with the human operating team. The team then helps the robot interpret its surroundings, and directs the robot to move or manipulate objects. The team repeatedly asks the robot to share its plan, adjusting its request and guidance until the robot provides a satisfactory answer. Once that happens the team permits the robot to proceed autonomously. "Effectively, we factor the problem of remotely-commanded dexterous manipulation into a human part and a robot part such that the whole system can perform the task end-to-end," Teller says.
In order to qualify for the VRC, MIT researchers first had to complete an initial round of competition in May. There they demonstrated that their software would allow a robot to successfully step across large gaps in terrain, and pick up and move a hand-held drill. During the VRC, the MIT team successfully operated a robot through a variety of maneuvers inside DARPA’s virtual simulator, such as walking across muddy, uneven terrain; dragging a fire hose and connecting it to a spigot; and opening a wall-mounted valve. The MIT team will now adapt its software and interface to a multi-million dollar humanoid robot, manufactured by Boston Dynamics and provided by DARPA, for use in the competition’s two remaining challenges, which are set for next December and December 2014.
The MIT team, based at CSAIL, is drawing on research that is underway in several departments and labs. For example, Teller’s own research group is developing methods that will make it possible for robots to interpret sensor data, allowing them to make sense of their surroundings and understand their movements. They are also creating human-robot interaction methods by which humans can help robots formulate and carry out plans. State-of-the-art, real-time planning and control methods that are being developed in Tedrake’s lab make it possible for humanoid robots to balance, walk and climb.
The MIT team includes faculty, research and technical staff, postdoctoral associates, PhD students, and undergraduate researchers from across MIT. Dr. Maurice Fallon of CSAIL leads the team’s perception and infrastructure efforts; Dr. Scott Kuindersma of CSAIL leads the team’s planning and control work; Dr. Sisir Karumanchi of the Lab for Manufacturing and Productivity leads development of the team’s manipulation capabilities; Dr. Toby Schneider of the MIT Center for Ocean Engineering in the Department of Mechanical Engineering has incorporated his methods for effective communication across low-bandwidth, high-latency network links, originally developed for underwater vehicles; and Dr. Matt Antone, an alumnus of the 2007 MIT DARPA Urban Challenge team and an MIT graduate, has returned to MIT to help with the DRC effort. The team also includes co-investigators Dr. Karl Iagnemma of the Laboratory for Manufacturing and Productivity, and Professor Julie Shah of the Department of Aeronautics and Astronautics.