DARPA code teams compete on same K9 robot

Multiple-personality droid dog; mad science indeed

DARPA, the Pentagon research bureau which likes to put the battiness back into battle-boffinry, is pressing ahead with its robot dog/packmule/mini-Imperial-Walker programme.

Partly-functional "BigDog" petrol-engined droid packmules have already been developed, but it seems the machines' controlling software isn't really up to dealing with rugged terrain.

Control software is one of the big problems in getting robots to match humans at tasks like climbing up mountains, marching over uneven ground, etc. You can build structures and actuators as strong and flexible as human limbs; you can generate as much power as the human body by using a fairly small petrol engine. But the damn machine still struggles to do what humans and/or animals can, because its software doesn't know how to handle its limbs and integrate what it does with what it sees.

Robotics researchers have struggled thus far to decide which software approaches work best. This has been at least in part because typically any given programme will use its own unique hardware, and someone else's code will have limited applicability.

DARPA is now looking to change all that - at least somewhat - by funding six different software teams to work on a common robot platform, the LittleDog. The LittleDog testbed was developed by Boston Dynamics, maker of the existing BigDog. It uses external motion-capture kit to monitor its own limb movements for now, but future versions will have internal feedback.

As you can see, the tiny K-9 isn't exactly greased lightning yet, but it's coming along. It actually reminds us more of an insect, somehow.

Max Lungarella, a robotics boffin at the University of Zurich, gave his thoughts on the DARPA LittleDog approach to NewScientist.

"What is really interesting about the whole project is the idea of a common research platform," he said. "A lot of research in robotics is done on platforms built ad-hoc."

DARPA's six software groups are at universities in America, including the usual suspects: MIT, Stanford, Carnegie Mellon, etc. Each group sends in its latest code monthly to DARPA, which uses it to control a LittleDog of its own across increasingly tough terrain. In this way, the Pentagon war-nerds can objectively assess which approaches are most effective, by trialling the six different controlling personalities of their robot dog against each other.

NewScientist reports that most of the groups have now moved away from stable walking, in which the LittleDog keeps three feet on the ground, to "dynamic" motion more like that which animals and people use.

According to Jerry Pratt of the Institute for Human and Machine Cognition (IHMC) in Florida, "walking becomes more like controlled falling".

For now, each university team has a duplicate of the test terrain used by the DARPA LittleDog. But next year the diminutive dog-droids will need to tackle previously unknown courses, identifying footholds and planning routes. DARPA will announce the winning team next year, and presumably will seek to use the knowledge gained in BigDog or similar future programmes. ®