Watch a Shape-Shifting Robot Prowl the Big, Bad World

Meet DyRET, a doglike machine that can lengthen its four legs on the fly. That’s not to creep out humans, but to help ramblin’ robots of all stripes.
robot dog
Courtesy of University of Oslo

Sure, evolution invented mammals that soar 200 feet through the air on giant flaps of skin and 3-foot-wide crabs that climb trees, but has it ever invented a four-legged animal with telescoping limbs? No, it has not. Biology can’t really work that way. But robots certainly can.

Meet the Dynamic Robot for Embodied Testing, aka DyRET, a machine that changes the length of its legs on the fly—not to creep out humans, but to help robots of all stripes not fall over so much. Writing today in the journal Nature Machine Intelligence, researchers in Norway and Australia describe how they got DyRET to learn how to lengthen or shorten its limbs to tackle different kinds of terrains. Then once they let the shape-shifting robot loose in the real world, it used that training to efficiently tread surfaces it had never seen before. (I.e., it managed to not collapse in a heap.)

“We can actually take the robot, bring it outside, and it will just start adapting,” says computer scientist Tønnes Nygaard of the University of Oslo and the Norwegian Defence Research Establishment, the lead author on the paper. “We saw that it was able to use the knowledge it previously learned.”

Walking animals don’t have extendable limbs because, first and foremost, it’s just not biologically possible. But it’s also not necessary. Thanks to millions of years of evolution honing our bodies, humans, cheetahs, and wolves all move with incredible agility, constantly scanning the ground ahead of us for obstacles as we run.

Go go gadget limbs…

Courtesy of University of Oslo

Robots, on the other hand, need some assistance. Even a super-sophisticated machine like the Boston Dynamics robot dog Spot has trouble navigating complex terrain. Giving robots telescoping legs both improves their stability as they move across different surfaces and boosts their energy efficiency. Stumbling around eats up a lot of battery power, and a flailing robot could hurt itself or nearby humans. “I think it's a particularly good idea to have a body that is tunable,” says Francisco Valero-Cuevas, an engineer at the University of Southern California who develops quadrupedal robots but wasn’t involved in this new research. “That's what's happening here. A tunable body makes for a more versatile robot.”

Nygaard and his colleagues schooled DyRET by first literally building it experimental sandboxes. In the lab, they filled long boxes with concrete, gravel, and sand, representing a range of different terrains the bot might find in the real world. Concrete is the easy one—nice and flat and predictable. Stepping in sand is much more uncertain, as with each step the robot’s legs would sink in unique ways. Gravel is a physically tough surface, like concrete, but it’s also unpredictable, as the rocks can shift, complicating DyRET’s footfalls. “By having the three terrain examples, with different hardness and roughness, you get a pretty good representation of a sort of general interaction between the morphology, or the body, and the environment,” says Nygaard.

Courtesy of University of Oslo

That morphology is quadrupedal, so DyRET moves like a dog or cat. Really, the robot is more or less just four legs with a handle on top for the researchers to grab. The robot’s legs can extend up to 6 inches total, but in two places: at the “femur” above the knee and the “tibia” below it. This gives the machine the capability of setting sections of its legs at different lengths. For example, it can telescope its limbs to have longer femurs and shorter tibias, or vice versa. The researchers could tweak these configurations, set DyRET loose on each terrain, and calculate how efficient each one was.

More specifically, they were looking at “cost of transport” as an efficiency measurement, the same metric that biologists use when looking at animal movement. Basically, it’s how much energy a creature or robot expends to convey itself, and how fast it moves. Stability while walking is inherently coded into that, which is of course important for an expensive robot like DyRET. “The more energy you expend not moving forward is energy typically spent being unstable,” says Nygaard. “So the less energy you spend moving forward, the more stable you inherently are.”

The researchers measured this energy expenditure in the motors in the robot’s joints and also used cameras to monitor its movement. The robot also had its own depth-sensing camera, which it used to characterize the roughness of a surface; for example, to observe that concrete is much smoother than gravel. The machine could even dip its toes in the water, so to speak: Force sensors on its feet gave it information about how much softer the sand was than concrete. Together, the camera and force sensors gave DyRET a complex picture of what it was walking on and how efficiently it was doing so.

Courtesy of University of Oslo

The researchers found that when walking across concrete, the shape-shifting robot was most efficient when it had longer legs. In sand, it moved efficiently with any femur length, as long as the tibia were short. On gravel, DyRET also excelled with shorter limbs overall, which makes sense: A lower center of gravity would give the robot better stability as it clambers over tiny rocks. Generally speaking, shorter legs allow the robot to apply more force to get a grip in looser material, while longer legs increase speed for walking across smoother material. (Above, you can see the robot lower itself when it detects that it's transitioning from concrete to gravel.)

All this training gave the robot prior knowledge of how best to configure its limbs for a particular surface. So when the researchers then took DyRET outside onto novel terrain, the robot could eyeball the ground with its camera and sense the give beneath its feet with the force sensors. Comparing this data with previous information about how concrete looks and feels, the robot then knew how to walk across a road—it made its legs longer overall for longer, more efficient strides. It didn’t need to worry about shortening its legs to lower its center of gravity, as it would when dealing with gravel, because it could see and feel that the surface was smooth and stable.

Courtesy of University of Oslo

DyRET could even tackle grass, a dramatically different surface than anything it had traipsed across in the lab. Its performance was iffy, at first. “It didn't really know what to do,” says Nygaard. “But then quite quick, it was able to sort of learn which body shapes perform better, and therefore adapt to this new environment as well.”

This isn’t a typical way to get a robot to learn to walk. As machine learning techniques have gotten more sophisticated over the past decade or so, roboticists have instead been training machines in simulation. That is, you train the software that controls the robot in a virtual world, where the simulated robot can make thousands of walking attempts, learning by trial and error. The system penalizes mistakes and rewards successful maneuvers until the virtual robot learns optimal behaviors, a technique known as reinforcement learning. Roboticists can then port that knowledge into the robot in the real world, and voilà, a walking machine.

Image may contain: Construction Crane
Everything you wanted to know about soft, hard, and nonmurderous automatons.

Except—not so voilà. This technique suffers from the “sim-to-real” problem: There’s just no way to perfectly simulate the complexities of the physical world in a virtual one, so the knowledge gained through simulation isn’t always square with the real world. That means the actual robot can wind up with a fuzzy understanding of its surroundings. Think of how well you’d get along if you woke up tomorrow and suddenly friction doesn’t work like you expect.

What these researchers have done with DyRET, by contrast, is simply to train the robot in the real world. That comes with its own challenges, of course: The shape-shifting machine learns much slower and could potentially get hurt. But the robot is also better equipped to deal with the absolute chaos of real surfaces and forces. “Differences in the terrain, and so on—like the roughness—these things are much harder to simulate than say, the high level of how you should walk, like trajectory,” says University of Oslo computer scientist Kyrre Glette, coauthor on the new paper.

Not only does DyRET have to adapt to different terrains, but to differences within those terrains. Grassy dirt, for instance, may be soggy or dry. The robot may hit a rock or a sprinkler, the kind of surprise that would trip up a robot trained in the simplified world of a simulation. With more and more real-world training, on the hand, DyRET can better prepare to tackle such obstacles without stumbling over them.

To be sure, this is early research: DyRET’s movement is still slow and stilted, especially compared to an advanced quadrupedal robot like Spot. Also, it can take up to 90 seconds to fully extend or contract the robot’s legs. But the researchers hope to both improve DyRET’s hardware and the underlying algorithms, perhaps one day getting to the point where other shape-shifting robots can adopt the same system. In fact, the larger idea generally in robotics labs is to get hardware and software to work more in concert—to make the machines better at sensing terrain and adapting their bodies and behavior to it. “This is a great recent example of how the interaction between the brain and the body is a very fruitful avenue,” says Valero-Cuevas. “That's only been recently catching on in robotics.”

And the robots will only get weirder from here. Imagine an eight-legged robot that can not only telescope its limbs, but choose when to use each of them. It might walk two-legged on flat surfaces, as humans do. “If the terrain gets steeper, at some point, you start scrambling on all fours,” says Valero-Cuevas. The steeper it gets, the more limbs the robot would activate to guarantee purchase on the terrain. “But when they're not needed, they could just fold away, and you're a very fast biped,” he says.

Beat that, evolution.


More Great WIRED Stories