Logo: University of Southern California

Teaching a RoboDog New Tricks

Mechanical quadruped software tested on difficult terrain in DARPA competitive trials
October 23, 2009 —

Stefan Schaal's Computational Learning and Motor Control Lab's quadruped robot program was the only one of four competitors in DARPA tests to maintain its target speed in five out of six runs.

DARPA Doggie
DARPA Little Dog climbing. To watch a narrated video of the dog in action, click on the image.
Schaal's was among six teams in the US to be awarded a DARPA Little Dog grant. The goal of the Learning Locomotion project is to program a small dog-like four-footed robot to find its way at a speed of more than 7.2 cm/sec over six different sets of obstacles and rough terrain using machine learning techniques.

More generally, according to the program's vision: "Large, irregular obstacles such as urban rubble, rock fields, and fallen logs present minor challenges to dismounted forces, slowing but not stopping them. These Slow-Go areas for dismounted forces are No-Go areas for today's small unmanned vehicles, limiting their effectiveness on the battlefield. Enabling future unmanned vehicles to traverse large, irregular obstacles will allow robots to better contribute to military operations."

“Of course, rough terrain can also be found in many daily life scenarios,” added Schaal. “For instance, a disaster site, a kids playground, or even a toy-cluttered room in a daycare center is currently not suited for existing robots.”

USC competed with CMU, IHMC (Florida), MIT, Stanford, and UPenn in monthly tests at on terrains the teams had never seen before a government site. 

The test was of the control software.The Little Dogs were standard and identical; the experimental setup was identical between the teams' labs and the test site so that it was possible to exactly compare pro-gramming approaches. After three years of research, only four teams — including USC — could cope with DARPA's final test scenarios..

USC stood out for robust general performance. The Viterbi School team made it its goal to be above the target speed of on all boards, while some teams opted to achieve high aver-age speed by “super-tuning” some behaviors

From left: Jonas Buchli,  Mrinal Kalakrishnan, Robodog, Stefan Schaal.
at the cost of substandard speed on some other boards.

USC was the only team exceeding the speed goals on 5 out of 6 test boards, thus demonstrating a level of robustness and performance that outperformed the final government program guidelines. In the final test, its dog achieved a total average speed of 8.7 cm/second.

The Phase 3 team of USC was composed of Mrinal Kalakrishnan, the lead RA in the project, Peter Pastor, Michael Mistry, and Jonas Buchli, the lead postdoc. “It was not one single component in our research that made the difference.” says Kalakrishnan, who specializes in real-time machine learning. “Rather a team effort where each team member contributed with well researched and tested components”.

Pastor worked on high performance behaviors including jumping and sliding. Only with these moves was it possible to reach the speed metrics for especially difficult terrains. Mistry developed theory and algorithms of novel model-based control algorithms that increased the accuracy of the dog’s movements. Buchli investigated very fast walking gaits that are robust towards slipping and other unforeseen disturbances.

“Compliant locomotion was one of the key features of our work” says Buchli. Compliant locomotion means that the robot gently “gives in” to disturbances, but continuously predicts and registers force feedback from the terrain at its feet which allow it readjust balance control accordingly. Such perception and control will also allow for safer human-robot interaction in the future.

Kalakrishnan designed the crucial foothold learning component in which the human user gives a demonstration of the best footholds in a given terrain. Using machine learning, the robot then generalizes this information to other terrains to navigate them without specific instructions. This research will ultimately allow robots to be taught by humans and then autonomously improve their behaviors by themselves.

Applications may include assistive robots that work in normal human environments, disaster area rescue, and “smart” prosthetics. Buchli says that “the dream is to take legged locomotion outside the predictable lab--it’s just not useful yet. Walking is a very misunderstood thing.”

Michael Mistry, left, and Peter Pastor
USC and other universities’ success proved that autonomous legged locomotion had potential that would have previously been considered impossible.

Schaal, a professor in the Viterbi School's Department of Computer Science, believes that performance in projects such as DARPA Little Dog confirms USC Viterbi among the top tier robotic programs in the U.S,  "In terms of robustness and fulfilling the government specs, USC did very well," he said.

He and his researchers are now awaiting the advent the full body humanoid robot that will arrive at USC in the next year in collaboration between the USC Computational Learning and Motor Control Lab and research counterparts in Japan.