DURUS Brings Human-Like Gait (and Fancy Shoes) to Hyper-Efficient Robots

A slick walking gait makes the DURUS humanoid robot even more efficient

4 min read

Evan Ackerman is IEEE Spectrum’s robotics editor.

DURUS humanoid robot from Georgia Tech
Image: AMBER Lab/Georgia Tech

In the middle of the DRC Finals last year, SRI’s DURUS robot slowly and steadily spent over two and a half hours walking 2 kilometers on a single battery charge. This was a Big Deal: DARPA had recognized from the beginning that the original version of ATLAS was horrendously impractical (at least in terms of locomotion), so they funded two different teams, one from SRI and one from Sandia, to design a humanoid robot that could walk 20x more efficiently. SRI’s DURUS came very, very close to this goal, achieving a cost of transport of just 1.5 through an innovative combination of hardware, software, and especially gait control.

The guy whose job it is to play with this robot is Professor Aaron Ames, who spent much of the last year moving his Advanced Mechanical Bipedal Experimental Robotics Lab from Texas A&M to Georgia Tech, which is why we haven’t heard anything exciting about DURUS since the DRC. 
It sounds like they just got everything up and running a few months ago, and they’re now ready to share an impressive new behavior: DURUS can now walk just like a human, while wearing normal (and stylish) human shoes.

You may remember that back at the DRC Finals, we asked Ames what he was hoping to do next with DURUS. At the top of his list (right after “better better better” and “faster faster faster”) was more natural foot behavior, including human-like heel-toe contacts. Check out this swagger:

The side-by-side (or rather, back-to-front) comparison at the end really drives home how closely DURUS’ gait matches a natural human walking gait. The only other human-scale robot that we’ve seen walk anything like this is Boston Dynamics’ PETMAN:

Not bad, but there’s the small matter of that not-so-subtle umbilical cord that’s providing off-board power to PETMAN. Massive amounts of off-board power. Like, an order of magnitude more power than it takes DURUS to move. To be fair, PETMAN was not designed to be efficient, and it can also do some tricks that DURUS can’t, like withstanding shoves and having arms and a head.

“The same technologies underlying DURUS . . . will translate to prostheses and exoskeletons to give mobility-impaired humans natural and efficient locomotion capabilities.”

DURUS has way better shoes, though, and the fact that you can run it for extended periods without being leashed to anything is hugely important. If we ever want robots to do practical things for us outside of a lab, they’re going to have to be capable of self-contained operation for a useful amount of time. 


DURUS is tethered for safety, but not for power. It’s walking with a cost of transport of 1.4, down from the 1.5 that we saw at the DRC Finals, thanks to this new human-like gait. “​The key feature of human walking is the way that the foot rolls during a step; that is, heel-strike followed by toe push-off,” Ames told us. Most humanoid robots, even the fancy and expensive ones like ATLAS, ASIMO, HUBO, and HRP, don’t try to do this. Instead, they walk on big, flat feet that provide a large contact area with the ground to maximize stability, while adopting a weird and very robot-y crouching gait to try to keep their center of mass over those giant feet as much as possible, since otherwise, they tend to fall over.

What’s unique about Georgia Tech’s approach with DURUS is that the methods used to develop it are based around a generalized mathematical framework that doesn’t demand any constraints or assumptions like the static stability offered by flat feet. Without these restrictions, you can simply add your own requirements to the framework, like a human-like heel-toe gait, and the robot will work it in. With DURUS, this resulted in more dynamic and natural looking gaits that boosted both speed and efficiency, as Ames explains:

​“In many ways, these results were a natural extension of what we unveiled at the DRC. ​The mathematical framework is the same, but we were able to extend it to more advanced behaviors. We have also increased our understanding (both mathematically and practically) on how to stabilize these much more dynamic and complex gaits; for example, the walking speed of this walking is about twice the speed of the gait at the DRC which requires everything to be that much better. But, in the end, the fact that this is the same fundamental mathematical framework shows the extensibility.” 

DURUS’ efficiency is a combination of both software and hardware. On the hardware side, it’s got those hefty ankle springs to help store and release energy with every step, along with feet designed to roll smoothly as they push off of the ground. On a fundamental level, though, these are just elements that can be leveraged by the mathematical framework that drives the motion of the robot, meaning that you can apply the same framework to robots with different hardware. And it’s not just humanoid robots like DURUS that will benefit, as Ames explains:

“The specific goal of this project is to create a universal control paradigm that is applicable ​to humanoid robots along with prostheses and exoskeletons. This, again, supports the development of ‘human-like’ walking gaits on humanoid robots. The same technologies underlying DURUS . . . will translate to prostheses and exoskeletons to give mobility-impaired humans natural and efficient locomotion capabilities.”

The AMBER Lab has funding from the National Science Foundation’s National Robotics Initiative to make this happen.


DURUS, meanwhile, will be kept busy learning new dynamic behaviors, Ames tells us:

“The ability to display human-like walking on a humanoid robot in the context of natural feet behavior is only an indication of what is possible with this framework. We can plan in much more advanced behaviors—one example is running. But going beyond that, we will be able to handle unknown and rough terrain, stairs, and other complex scenarios. This is all future work, and will be an essential component in bringing robots into the wild.”

The planar (and tethered) version of DURUS has already demonstrated human-like running behavior, and we can’t wait to see this kind of thing on an unsupported robot. And that’s really the key here—using the unified control and design framework that AMBER Lab is developing, you generalize different behaviors across different hardware, making it much more practical to develop everything from big walking robots to lower body assistive exoskeletons and prosthetic systems. DURUS’ job is to enable this development path through exploring more robust locomotion and a broader suite of walking behaviors, which, says Ames, “are the first steps toward systematically understand how to enable humanoid robots to walk among us.”

[ AMBER Lab ]

The Conversation (0)