While disaster robots were making their way through the DARPA Robotics Challenge courses, over in the exhibit area outside, there was another competition taking place: an endurance challenge, also sponsored by DARPA, where robots from Sandia National Labs and SRI International slowly walked on treadmills with the goal of demonstrating how ultra-efficient they could be.
What does ultra-efficient mean in the context of walking robots? Think humanoid walking that’s 20 to 30 times more efficient than than Boston Dynamics’ ATLAS. A full size humanoid robot with that level of efficiency would able to operate for anywhere from 4 to 6 hours on a single charge.
The robot that SRI displayed at the DRC is called DURUS (the root there being “durable”), which was developed under a (super duper secret, apparently) DARPA contract beginning in 2013.
That year was also when Boston Dynamics unveiled their ATLAS robot, and that’s not a coincidence: DARPA saw ATLAS, immediately realized that power requirements made ATLAS an entirely impractical platform for real-world use, and started funding a program to develop ultra-efficient actuators. This was expanded into an efficient mobility challenge, with the goal of building a humanoid architecture that was similar to the ATLAS robot (with the same kinematics), except 20x to 30x more efficient.
The culmination of all this was an efficiency walk-off between SRI’s DURUS, and Sandia’s robot. Here’s Professor Aaron Ames (until recently at Texas A&M and now at Georgia Tech) describing what makes DURUS so awesome:
The competition involved giving DURUS a fully charged battery, setting it walking on a treadmill (tethered for safety but otherwise self-contained), and seeing how long it could walk before it ran the battery dry. We were there to see its final steps, and in 2 hours, 35 minutes, 43 seconds, DURUS walked 2.05 kilometers. While walking, it used about 350 watts of power, giving it an average cost of transport of about 1.5.
Cost of transport is a convenient measurement of the efficiency of moving something somewhere. ATLAS has a cost of transport of something like 20. MIT’s cheetah quadruped is at about 0.5. A human is down around 0.2. SRI estimates that with some additional tuning, they’ll be able to max DURUS out at just under 1, and with its onboard 2.2 kWh lithium-polymer battery, the 80 kg robot should be able to walk 10 km (!).
It’s important to note that DURUS isn’t taking advantage of some new power source technology with some super high energy density, or novel ultra-lightweight construction, because that’s not what this challenge is about. The goal was to find efficiency in actuation and robotic walking motion. To that end, the SRI team looked at every single location where efficiency is lost in a typical robotic system: motors, mechanical transmissions, motor controllers, electrical wiring, kinematics, dynamics, walking gait… Everything. The efficiency of DURUS is a combination of improvements in all of these areas, but there are definitely a few highlights.
SRI identified the combination of motors, motor controllers, and transmissions as a big source of inefficiency. The transmissions were especially bad, so SRI decided to reinvent them. They wouldn’t tell us exactly what they did, except that “everywhere there’s sliding friction [like in a traditional gearbox], we tried to turn it into rolling friction instead, leading to an improvement in efficiency that is a few orders of magnitude better than traditional gearboxes.” Whatever they came up with manages to be about 97 percent efficient, enabling low-cost electric motors to make efficient actuators.
But the biggest improvement embodied in DURUS is by far its walking gait. If you spent any time watching the DRC, you probably noticed that all the robots walked like, you know, robots. The stereotypical quasi-static robotic gait is a side effect of the robot trying to keep its center of mass balanced above its feet, which is the way to go if you’re worried about falling over all the time. Humans don’t walk like this. Instead, we walk by falling forward intentionally, constantly catching ourselves by taking steps. DURUS walks the same way, thanks to software controls that provide a realistic level of confidence that the robot can walk continuously without falling over.
DURUS is a research platform for DARPA that’s basically the lower half of a full humanoid that SRI is developing, which will (eventually) be named PROXI [right].
PROXI will meld SRI’s experience with arms and hands with the efficiency of DURUS to create a low cost, high performance, electric humanoid that can walk for 8 hours. And now that the DARPA funding for this project has run out (it ended with the DRC Finals), SRI is thinking about commercial production. Rich Mahoney, the director of SRI’s robotics program, explains what we have to look forward to over the next few years:
“We don’t believe that there’s a platform [that exists right now] that has the kind of components, performance, and dynamic response that PROXI will have. Hopefully we’ll see a path where initially some research groups will start with them, and then in 3-5 years, if we get the volume, this is a robot that could be on sale for under $100,000. And even potentially in the $50,000 range, with any kind of reasonable volume. We have something that can open up a market: the platforms are getting ready to emerge that will enable the next generation of robot applications, and I think this platform will be one of those.”
[ SRI ]