In early June, robotics professor Jessy Grizzle and his students took the blind, two-legged robot MARLO out to the Wave Field, an earthen art installation, for the first time. The tech news saw the videos the team posted and dubbed the early attempts “drunken”—MARLO couldn’t make it even halfway down the troughs between the bigger waves, and forget trying to handle the large mounds. But at that moment in mid-June, it was an accomplishment just to be out there on the field that IEEE Spectrum called “a walking robot nightmare.”
MARLO fell repeatedly, but the team quickly figured out why. MARLO wasn’t lifting its feet high enough for the steep slopes. When the robot tried to prevent a fall by putting its foot out to the side, it failed when that foot stopped short, ramming into the side of a mound.
Xingye “Dennis” Da, a doctoral student in mechanical engineering wanted to see MARLO walk on the Wave Field before he left the country for a month.
“It is heartbreaking to watch the episodes of falling on the Wave Field over and over again. We treat MARLO as the fourth member in the lab and she deserves to have a better controller to walk safely,” said Da.
So over the next few weeks, he analyzed the failure patterns of all of those falls and produced a new controller that could prevent them. Meanwhile, the team worked long hours to upgrade MARLO’s hardware to meet the demands of the difficult terrain—with the help of MARLO’s makers, Jonathan Hurst’s group at Oregon State University.
Previously, Da’s work had focused on the controller for MARLO’s forward and backward motion. He produced a library of 15 gaits for different ground heights and walking speeds and designed an algorithm that could blend them on the fly. But it was separate from the controller for side-to-side motion. To succeed on the Wave Field, MARLO needed to integrate information from both controllers.
Da hand-coded this new “super-controller” to estimate changes in slope with each impact of the robot’s feet on the ground, using the angles of MARLO’s joints. It then extrapolated how high the feet would need to lift in sidestep maneuvers.
For initial testing, Da developed a computer simulation of the Wave Field for a digital MARLO to practice on.
“It’s a slightly tamer version of the Wave Field, but our previous controller could not have stood up to that simulator,” said Grizzle, the Elmer G. Gilbert Distinguished University Professor and the Jerry W. and Carol L. Levin Professor of Engineering.
The night of Friday July 22, following a party to celebrate lab mate Brent Griffin’s successful thesis defense (PhD EESys ’16), Da went back to the lab with Omar Harib, a doctoral student in electrical and computer engineering, to run a final test. They replicated a single Wave Field mound with foam, plywood and astroturf and marched MARLO up and down it, over and over. At first, Da said MARLO couldn’t even walk in place. They struggled with a bug in a single line of code.
Around 2:30 AM, they called it a night. They were back at the lab at 9 AM on Saturday to try the real Wave Field. MARLO is only allowed outside on evenings and weekends, per college policy.
“To be honest, my expectations were low, but this was our last chance until the fall,” said Grizzle.
“Right off the bat, we walked all the way down one of those troughs…We just couldn’t believe it.”
Grizzle’s team uses MARLO to test control algorithms that can be applied and adapted to any robot designed for two-legged walking, including motorized prosthetics and Georgia Tech’s DURUS. The improved agility will help robots adjust to rough terrain in a human-like way, without the need to watch the ground.
They pushed MARLO to its limits by climbing one of the knolls. The experiment ended when MARLO stumbled down the other side. “We completely destroyed a leg,” said Grizzle. “There was twisted metal and shredded aluminum.”
Three days later, Da was on the plane back to China. Ross Hartley, a doctoral student in robotics, and Harib put MARLO back together.
Impressive as the feat may be, the algorithms aren’t quite ready to be shared with other roboticists. Da took a shortcut and coded the new super-controller specifically for MARLO. Once he is back, he and Hartley will generalize the code.
Jessy Grizzle is also a professor of electrical engineering and computer science and a professor of mechanical engineering.