The Clumsy Quest to Perfect the Walking Robot

Jerry Pratt saw sky on the video monitor and knew he'd fouled up. The sky—southern California blue as usual on a June afternoon in Pomona, about 30 miles east of Los Angeles—wasn't the problem. The problem was that there was only one reason to see sky on a monitor connected to a camera connected to the head of a very expensive, very sophisticated humanoid robot. Instead of stepping nimbly onto a small pile of cinder blocks, the robot, nicknamed Running Man, had fallen flat on its back.

Pratt did not see the fall happen, but the crowd of roboticists, journalists and spectators gathered around the course at the 2015 DARPA Robotics Challenge (DRC) did. Pratt and his collaborators from the Florida Institute for Human & Machine Cognition (IHMC) were here competing against 24 other teams to win a $2-million prize. And for the moment, Running Man lay frozen, its right leg stuck skyward like a postpratfall comedy actor waiting to hear the director shout, “Cut!” Then gravity reasserted itself, and the robot's hips and torso lolled to the side, the dead weight of its legs sagging slowly to the pavement. Its long arms lay flat and splayed, snow-angel-style.

This is not what Pratt and his partners were going for. They and the other teams, including competitors from top robotics laboratories such as ones at Carnegie Mellon University and the Massachusetts Institute of Technology, had come to show that their robots could do simple things most able-bodied humans take for granted—like opening doors, driving motor vehicles, manipulating hand tools and walking around on two legs. In 60 minutes or less, DRC contestants had to drive, steer and exit a small Jeep-like vehicle, open a closed door and enter a building, clear debris out of a hallway or traverse a pile of jumbled cinder blocks, pick up a power tool and cut through a panel of drywall with it, turn a large metal valve and ascend a short staircase. Many of the robots managed at least several of these feats, but they also fell down. A lot. The most enduring footage of the event is destined to be a highlight reel of robots toppling over like inebriated college students; on YouTube, it has been viewed 1.8 million times and counting.

On supporting science journalism

If you're enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.

Six months after the competition and back in his lab in Pensacola, Fla., Pratt recalled the event and offered assurance to anyone worried about a humanoid-robot uprising. “Walking,” he says, “is hard.”

Baby Steps

Walking is hard—just observe a child under the age of two or talk to anyone undergoing physical therapy to regain the skill after an injury. But why is it hard? After all, our species has been bipedal for hundreds of thousands of years; other bipeds such as ostriches have been walking for millions more. “One has an intuition that if babies walk, it must be easy,” says Andy Ruina, a professor of mechanical engineering at Cornell University who has been studying legged locomotion and designing walking robots since 1992. “But babies can do all kinds of things that we still don't understand.”

What babies struggle to master about walking could be summed up in one word: agility. Stepping, balancing, maintaining momentum, correcting for errors, adapting to terrain—each of these complex behaviors is necessary but insufficient for bipedal locomotion. Degrade any one of them even slightly, and the smoothly integrated act of walking that most healthy adults take for granted quickly becomes clumsy, fragile and enervating.

Biological agility exhibits the opposite characteristics. First, it is controlled: we use our senses to find our footing with confidence and reliability. Second, it is robust: most of the time we can accommodate surprises and recover from errors. Third, walking is efficient: it does not require maladaptive amounts of time, energy or attention to perform routinely. In other words, an organism ought to be able to walk and chew gum at the same time.

Adult humans pull off this trifecta with an ease honed by millions of years of evolution (not to mention several years of continuous practice in the early stages of life). We learn control and balance using sight, touch and proprioception. Our reflexes ensure that we do not pitch into the dirt every time we encounter an unexpected pebble, and strong bones surrounded by flexible tissue protect against most spills. Finally, our every step is a symphony of mechanical and computational efficiency: our muscles and tendons can passively absorb impacts one instant and actively generate propulsion the next, and our spinal cords maintain periodic motor patterns that keep our legs moving in the right direction while our brains tend to more important business.

This, then, is what makes robotic walking “hard”: no bipedal robot has yet been engineered to combine control, robustness and efficiency as well as humans—or chickens, for that matter—do. Honda's famous, astronaut-faced ASIMO meticulously calculates the force, trajectory and momentum required for each step: a control-focused approach. Boston Dynamics, whose viral videos show its next-generation Atlas humanoid hiking through a snowy forest and picking itself back up after falling down, emphasizes robustness: speed and balance over planning and precision. (Running Man and several other DRC competitors were modified from Atlas robots.) Aaron D. Ames, a robotics researcher at the Georgia Institute of Technology who is working on a headless, armless biped called DURUS, specifies every possible degree of freedom in the robot's body with dense mathematical equations that would fill hundreds of pages each if written out. Jonathan Hurst, a mechanical engineer at Oregon State University, built a relatively simple robot called ATRIAS based on a generic physics model that also describes the behavior of ground-running birds. Despite their differing approaches, both Ames and Hurst are interested in the same thing: efficiency. Hybrid strategies exist, too. Pratt, whose Running Man won second place in the DRC, used a method called capture point to blend ASIMO-like control with Boston Dynamics–style stability. Each approach has its advantages, yet none can emulate the efficiency, flexibility, speed and precision of an adult human walking.

It is tempting to say that engineers shouldn't bother trying. After all, Jun Ho Oh, a leading roboticist at the Korea Advanced Institute of Science and Technology (KAIST), won the $2-million DRC first prize not by outwalking the competition but by avoiding legged locomotion as much as possible. (The robot was equipped with wheels on its knees and feet, allowing it to roll in a stable kneeling position through much of the course.) And the Wright brothers didn't invent the airplane by slavishly mimicking the way bird wings flap.

Yet there are valid reasons that the desire to build walking robots endures. The most obvious application is as old as the legend of the Golem: a stronger, better version of a human body could accomplish tasks deemed too difficult, dangerous or tedious for people to risk by using their own bodies. The DARPA Robotics Challenge itself was designed to mimic the meltdown at Japan's Fukushima Daiichi nuclear power plant in 2011. That disaster might have been mitigated if robots had been capable of driving into the plant, navigating a few staircases or debris-filled corridors, and turning some valves or switches. Disaster response would just be one application. Office telepresence and home assistance, package logistics and delivery, security patrols and safety monitoring, and resource exploration and extraction could all conceivably be augmented or automated by humanoid robots. “I know of no existing biological or mechanical form that is better suited for terrestrial locomotion than a humanoid,” Pratt says.

The indirect benefits of robotic bipedalism could be significant as well. Ames says that building robotic systems that capture the full range of human locomotion will help us understand walking itself. “If you can make robots walk like people, you can help a lot of people who can't walk,” he says.

The promise of a bipedal humanoid mirrors the promise of artificial intelligence: that of unbounded versatility. If an AI is the ultimate thinking machine, a humanoid robot could become the ultimate “doing machine”—a truly general-purpose meta tool, capable of getting to and performing in unpredictable environments while taking advantage of all the useful devices we have already invented. “The only reason to have a humanoid robot is to have a general-purpose robot,” Oh says. “Bipedal robots are not necessary everywhere. But somewhere it will be necessary. We are preparing for those situations.”

Complete Control

Pratt's IHMC robotics laboratory looks like an indie maker space crossed with a small software start-up. Two young researchers ride caster boards around an open bullpen of standing desks, firing Nerf guns at each other. A cluttered, hangarlike work space houses a replica of the DRC course and a large metal gantry. Dangling from the gantry like a side of beef is Running Man: limbs slack, slab-shaped feet tipped downward, toes just barely touching the concrete floor. John Carff, the lab's most experienced robot operator, launches Running Man's calibration routine. Still dangling, the robot begins raising its arms and legs into a precise pose, like Leonardo da Vinci's Vitruvian Man.

Leaning against a pillar next to the gantry is a long length of white pipe with a miniature red Everlast boxing glove taped to one end. It represents a minimum safe distance one should keep from Running Man when the robot is active. “We have an official ‘10-foot pole,’” Pratt explains, as a motorized pulley lowers Running Man's feet to the floor. A thick electric cable delivers 10 kilowatts of power to the 386-pound humanoid's hydraulic actuators. “That's about the strength of 12 horses coursing through its veins,” Pratt says. “If something went wrong and it whacked you in the face, it could kill you.”

As Pratt and Carff put Running Man through some of the paces it made at the DRC, though—balancing on one foot, walking a few steps toward a cinder block, stepping onto the block and off—it seems at once imposing and unnervingly fragile. Its hulking torso, encased in a padded metal roll cage, balances over tapered legs bent at the knee, giving the robot a burdened, top-heavy appearance. Its push-recovery software is not installed at the moment, so a decent shove could tip it over (although the safety harness attached to the gantry would catch it). It moves with the deliberate pace and shuffling gait of an elderly person guiding a walker across a traffic intersection. But its steps are sure: the safety harness slackens visibly as Running Man pushes its considerable bulk atop the cinder blocks.

The method Pratt uses to plan Running Man's steps, known as capture point, refers to the position on the ground that a biped's foot must reach to stop its body from falling over. When someone is striding quickly or running, the capture point for each step does not need to be determined as precisely in advance, because the biped spends a relatively short time balancing before taking the next step. When a person is walking slowly or stepping over uneven terrain, however, “the placement of each step is more critical,” Pratt explains. “If you're off by a few inches, you're going to be staggering all over the place.”

Think of using stones to cross a creek without falling in. One approach would be to quickly “fall forward” step by step in roughly the right places on the stones to maintain your balance and trajectory. The other would be to move slowly and carefully, placing your foot in just the right spot to safely transfer your weight with each step.

According to Pratt, Running Man's real-time ability to sense its own position in space—accomplished via an inertial measurement unit in its pelvis and software that recomputes balance and orientation 1,000 times per second—far outpaces humans. But what humans have that Running Man does not are lightweight, flexible limbs capable of moving quickly enough to correct for errors or disturbances on the fly. Pratt describes a game he plays with his sons that makes this point clear. “We'll be walking down the street, and all of a sudden I'll yell, ‘Push recovery!’ and give 'em a shove,” he chuckles. If Pratt were to pull the same prank on Running Man, there is a decent chance that even with its sophisticated push-recovery programming and 12-horsepower hydraulic joints, the robot would simply crash to the ground.

Another advantage that human bodies have over robots is the ability to get back up after falling—or at the very least, to not shatter into pieces. “You're landing on big pieces of heavy metal, so it's hard to build something that can survive a fall,” Pratt says.

So if the bipedal robots at the DRC walked more like tense, frail old people than fearsome, agile Terminators, it is because their bodies forced them to do so. “The bottleneck [to engineering bipedal robots] isn't computational at all. It's the hardware,” Pratt says. “If we could build the robot with something that had the same properties as muscle”—a lightweight, energy-efficient actuator capable of behaving like a powerful motor one moment and a passive spring the next—“I think it could be really good.”

The Spring-Mass Model

Hurst's walking robot, ATRIAS, is blind as a bat and dumb as a rock. It does not even have a head—just a metal pole jutting out of its boxy black thorax, which Hurst and his graduate students use to guide the robot as it struts around Oregon State University's Graf Hall like a decapitated mechanical chicken. Despite these apparent deficiencies, however, ATRIAS can perform a surprisingly humanlike feat that no path-planning, capture-point-calculating biped at the DARPA competition could: it can trip over an unexpected obstacle and keep on walking as if nothing happened. Compared with the plodding footwork of most bipedal humanoids, ATRIAS is Gene Kelly performing “Singin' in the Rain.”

“We designed this robot as a scientific tool for one purpose: investigating the fundamental principles behind walking,” Hurst says. In other words, don't expect to see ATRIAS ostriching its way into any future disaster sites. But if his understanding of bipedal walking is correct, we may not have to wait for someone to invent artificial muscles before robots can walk with animal-like robustness and efficiency.

ATRIAS stands for “assume the robot is a sphere,” a physics in-joke that basically means “keep it simple, stupid.” Its behavior is based on a decades-old theory of legged locomotion called the spring-mass model. According to this model, all the variables needed to describe a walker made of bones, muscles and tendons can be abstracted to just two elements: the mass of a body attached at a single point to a massless (in the real world, as lightweight as possible) spring-equipped leg.

Click or tap to enlarge

The spring-mass model is little more than a computer-controlled pogo stick with a weight on top. But this model has informed the engineering of legged locomotion in robots for decades, most famously at the M.I.T. Leg Lab, where founder and principal investigator Marc Raibert conducted pathbreaking research in the late 1980s and early 1990s on hopping and running robots before leaving academia to found Boston Dynamics. (Hurst and Pratt also spent time at the Leg Lab before starting their own respective legged-robotics labs.)

The spring-mass model is important because it provides one of the foundations for an important feature of walking called dynamic stability. A dynamically stable walking robot maintains balance in the same way a human does: by catching itself midfall with each step. If a disturbance or mistake interrupts its stride, and the walker cannot correct its gait in time to support its center of mass, the walker will fall down. “A human's center of mass is about [three feet] off the ground, which means you have to swing your leg into place in less than one third of a second to avoid a significant fall,” Pratt explains.

Static stability, meanwhile, takes the opposite approach: rather than maintaining a state of controlled falling, it treats walking “as a perturbation of standing still,” Ruina says. The path and momentum of each step must be precisely calculated in advance so that the robot's center of mass stays continuously balanced at every point in its stride. Early bipedal humanoids used statically stable walking to make the robots' rigid limbs easier to control and allow them—in theory, at least—to freeze midstep at any point without falling over. Contemporary humanoid robots, including the ones competing at the DRC, still use a version of this approach called quasi-static stability, which requires similarly deliberate, flat-footed steps to maintain balance.

A quasi-static biped walker needs a lot of energy-guzzling actuators and computational power to control its stiff, bent-kneed gait—and it is still exceedingly sensitive to disturbances. But a dynamically stable biped based on the spring-mass model, like ATRIAS, off-loads much of this business to the naturally occurring physical interactions between its legs and the terrain. “If you're walking on rocky ground, you can swing a leg out, flop it down, and it'll just conform automatically to whatever it hits,” explains Pratt, who took advantage of similar dynamics while designing bipeds at the M.I.T. Leg Lab in the late 1990s.

When combined with strong hip motors and legs that can swing passively (without being pushed by motors), the spring-mass model can produce an efficient gait that is surprisingly resilient when disturbed. Hurst uses the phrase “animal-like performance” to describe the combination of energy savings and agility in ATRIAS's walk. Indeed, when he tracked the robot's movements and plotted the data over time, the resulting curve closely matched that of a human and several species of ground-walking birds.

Hurst says this correspondence implies that the physics he used to design ATRIAS's body and behavior might be identical to some of the principles underlying natural bipedalism. “We're not doing any biomimicry with ATRIAS,” he asserts. “Its legs look nothing like a chicken's or a human's. But the walking patterns we see are the same underneath. That tells me we're onto something—and it probably doesn't require faster actuators or more computation.”

Adapting While Walking

Oh's DRC-Hubo+ may have won the $2-million first prize at the DARPA Robotics Challenge, but a glance around Oh's lab proves that his is not an overnight success. Nestled in a bunkerlike workshop on KAIST's campus in Daejeon, South Korea, HuboLab is festooned with obsolete iterations of the humanoid robot that Oh has spent the past 15 years bootstrapping.

They hang from small gantries like suits that have gone out of style: the original Hubo design, a child-sized ASIMO facsimile that Oh cobbled together in 2004 using leftovers from his colleagues' research budgets after being refused funding by the South Korean government; a version that did receive funding following Oh's successful proof of concept, its gray outer shell now stripped away and its metal innards exposed like a robotic Body Worlds exhibit; a black, headless prototype of Hubo that Oh constructed to stress-test his latest designs for the rigors of DARPA's simulated disaster scenario. DRC-Hubo+ itself is more like a life-sized GoBots toy, with gleaming red and blue accents adorning its slim, geometric, brushed-aluminum body. And much like that same toy, Hubo's secret weapon is not brains or strength but an ability to transform its humanoid shape in surprising ways.

Oh carries himself with a jolly, slightly eccentric air. He is comfortable putting on a bit of a show—especially since receiving that $2-million check from DARPA. One day in early February, Oh and his graduate students performed Hubo demos for a visiting delegation of French technocrats, the president of KAIST and a Korean military official. The week before, he had accompanied Hubo to the World Economic Forum meeting in Davos, Switzerland.

Given Hubo's accolades and celebrity, one might assume that Oh has boundless faith in his robot's walking ability. Instead he jovially recounts how often Hubo fell in the run-up to the DRC—“about once a month, but mostly the damage wasn't that serious,” he says—and openly admits that his winning strategy depended on avoiding bipedal locomotion wherever possible. “If walking works 99 percent of the time in the lab, the 1 percent in reality is always where the problems are,” Oh says.

Oh originally intended Hubo to walk through the DRC course, just like Running Man. But after repeated difficulties during testing, Jungho Lee—a fellow roboticist and co-founder of Rainbow, a spin-off of HuboLab that commercializes the robot and its technology—convinced Oh to take a more conservative tack. Instead of staking their chances on Hubo's imperfect walking, Oh came up with a solution he calls “multimodal mobility.” Another phrase that comes to mind is “whatever works.”

Essentially Oh turned Hubo into a Transformer. On even ground, the robot folds into a kneeling position and drives around on wheels affixed to its knees and feet. Hubo's torso can also spin independently from its pelvis, which lets the robot twist itself into positions that can maximize its effectiveness in clever ways. For instance, when faced with the DRC's debris-filled hallway, Hubo did not waste time—or risk falling—by removing the obstructions by hand from an upright position. Instead it knelt on its wheels, rotated its upper body 180 degrees and used the flat bottoms of its feet—now facing “forward”—as a ram, pushing the debris out of its way as it rolled swiftly and safely ahead.

Oh's ingenuity produced a legged humanoid robot whose performance combines precision, robustness and efficiency while staying true to the letter of the DRC's rules. But what about its spirit? It depends on which roboticist you ask. “I didn't love that,” says Georgia Tech's Ames of Hubo's transformations. (His DURUS humanoid did not compete in the main challenge course, but it did win the Robot Endurance Test, a sideline competition for ultraefficient bipedal walking.)

Tony Stentz of Carnegie Mellon, whose third-place-winning CHIMP robot eschewed walking in favor of rolling on four legs equipped with tank treads, has a different opinion. “You have to look at the problem and come up with the best design to solve it, all factors considered,” he says. “If you just come out [to the DRC course] and say the solution must have a bipedal form, then I'd say you're greatly constraining your solution—and it's possible that you no longer have the optimal solution.”

Oh agrees, even though he is as bullish on the utility of humanoid bipeds as anyone else at the DRC. “If [humanoid] walking were perfect 100 percent of the time, we wouldn't need multimodal mobility,” he says. He shares Pratt's belief that hardware is what is holding humanoids back; he plans to devote the next two years to building up an understanding of actuators “from scratch.” Still, Oh adds, “I'm not going to wait for innovative actuators—so we have to rely on electric motors, hydraulics or pneumatics” to refine Hubo's effectiveness. If that means devising clever locomotive hacks to compensate for humanoids' imperfect performance on two legs, so be it.

Almost Human

In February, Boston Dynamics released a video of its new humanoid robot doing almost all of the things that the robots competing in the DRC had struggled or failed to do. The new robot—a redesigned version of the Atlas humanoid that several DRC teams had used—could approach a door, open it and walk through it at a brisk, humanlike pace. It marched down an uneven embankment and regained its balance even as its feet slipped to and fro on the snowy ground. It confidently lowered and raised its body from a squatting position while grasping a 10-pound weight. It fell down on its face—hard—without shattering or spraying fluid from a ruptured hydraulic vein (as one unlucky DRC competitor memorably did in 2015). And perhaps most impressively, it pushed itself onto its feet and stood back up.

The demonstration hit the humanoid robotics community like Deep Blue beating Garry Kasparov at chess. Ruina called it “a game changer.” Ames and Hurst, respectively, deemed it “spectacular” and “the real deal.” Pratt praised its “phenomenal” range of motion, especially “the way it can squat all the way down.” He added, “I can't even do that.” Still, none of them consider robust bipedal walking to be “solved” and not just because Boston Dynamics refuses to share the scientific or engineering details behind its creations. “This is the new state of the art,” Ames says. “What they presented is a solution, and theirs is better than most others, clearly. But it's not the solution.” (Boston Dynamics did not respond to repeated interview requests from Scientific American.)

For these researchers, the same questions still persist. How can mechanical actuators deliver powerful torque and exploit passive dynamics at the same time? What control algorithms will let a robot manage the difference between tiptoeing carefully up a staircase and striding swiftly up a mound of boulders? How will the engineering of the system scale up in efficiency and down in price? “There's no Moore's law for this,” Pratt says.

And so the work to solve bipedal walking continues. Hurst is working on a successor to ATRIAS that can already run, walk, steer itself and pick itself up off the ground in simulation. Ames plans to get DURUS off the treadmill and walking around Georgia Tech's campus sometime in 2017. Ames and Pratt are contributing to NASA's Valkyrie project, aimed at developing a humanoid robot to accompany astronauts to Mars; meanwhile Ruina is developing a biped called Tik-Tok that he claims will demonstrate humanlike efficiency and performance using cheap, off-the-shelf components.

“We had hopes of making a video like [Boston Dynamics's] within a year or two, so they did take that away,” Ruina admits. “For a minute I thought, ‘Shoot, now what am I going to do for the rest of my life?’ But then I thought of the Wright brothers. Their invention wasn't the end—it was the beginning. The theory of airplane dynamics came along afterward. Atlas is by far the most impressive biped anyone's ever made. But does it mean there's nothing left to do? No. This opens up a whole world of thought.