An Algorithm Helps Robots Fall Safely
Researchers at Georgia Tech have developed an algorithm to help humanoid robots hit the ground without breaking themselves.
Robots that provide assistance in homes, hospitals, and workplaces will need to be able to fall without damaging themselves.
At a major robotics competition held in June, several multi-million-dollar robots struggled to perform even simple tasks like climbing a flight of stairs; some even toppled over with what seemed like impeccable comic timing. But a few of these amusing pratfalls resulted in devastating damage to the robots’ instruments, motors, and other components.
Fortunately for such robot klutzes, researchers are exploring ways to enable robots to fall more gracefully and safely. The work will become important as robots are used in more complex environments, and as engineers experiment with machines that move around on legs rather than wheels.
Researchers at Georgia Tech took inspiration from the way people break a fall by sticking out an arm or a leg on the way down. “When you fall down, you try to dissipate energy,” says Karen Liu, a professor of computer science at Georgia Tech who carried out the work with her then graduate student Sehoon Ha, who now works at Disney Research Pittsburgh. “And every time you make contact with the ground, some of that energy is dissipated.”
Liu and Ha devised an algorithm that lets an unbalanced robot figure out how to contort its body so that it hits the ground with less force. The algorithm calculates how to create a number of contact points with the ground in order to disperse the momentum of the fall.
At a conference in Germany last month, the Georgia Tech pair described testing the algorithm with a small humanoid robot called BioloidGP, and in simulations of a large humanoid called Atlas. The latter was developed by a company called Boston Dynamics, which specializes in making advanced legged machines and is now owned by Google. Several teams involved in the June event used Atlas robots. At the event, called the DARPA Robotics Challenge, tele-operated robots raced to perform a series of tasks including driving a golf cart, opening a series of doors, and operating a power drill.
The DARPA event was primarily meant to simulate the problems a robot would encounter when assisting at a stricken nuclear plant, but it also highlighted the remaining challenges for robots aiming to work in just about any normal human environment (see “Why Robots—and Humans—Struggled with DARPA’s Challenge”).
Matt DeDonato, who led a team from Worchester Polytechnic Institute at the DARPA event, says that most participants were more focused on staying upright than working out better ways to fall, especially because each fall incurred a hefty time penalty. To minimize damage, the Atlas robot operated by his team, in collaboration with researchers from CMU, would power down its actuators and go limp when it detected a fall. But DeDonato, whose team managed to keep the robot upright throughout the DARPA event, says the area needs to be explored as more robots become commercialized. “You are guaranteed to fall over sometimes,” he says.
Marc Raibert, founder of Boston Dynamics, now part of Google, and a pioneer in legged robotics, says his team started thinking about how to protect a falling robot while developing a four-legged machine called BigDog. The first idea was to have the limbs seize up when a fall was detected. “That caused the limbs to act as long levers that apply large forces to the joints when the limbs strike the ground,” he says. “We actually broke some legs clean off the robot, so we reprogrammed BigDog to relax its joints during a fall. All the robots we build now do something like that when they detect that they have lost their balance.”
Lui says her group is also interested in devising ways for robots to avoid hurting people if they fall over. This might involve balancing in such a way that made falling toward a person more unlikely, she says.
The approach developed so far is limited, however, by the sensing capabilities and the computational power of most robots. For their experiments, the Georgia Tech team used an accelerometer in the head of the physical robot as well as external motion-capture cameras. Liu notes that the complexity of calculating how best to fall explains why so many animals, including humans, have a nervous system, which reacts automatically.
“That’s why we have reflexes,” Liu says. “We are thinking of building something like a nervous system for robots.”
Become an MIT Technology Review Insider for in-depth analysis and unparalleled perspective.Subscribe today