Editor's note: Legged robots have the ability to follow troops on long journeys across extremely difficult terrain. In our series on legged robotics, Scientific American Online explores the challenges such technology poses as well as two DARPA projects—BigDog and LittleDog—that have shown great promise.

The Defense Advanced Research Projects Agency (DARPA) has for years explored the possibility of using legged robots to carry troop supplies where wheeled robots dare not tread (particularly through narrow mountain passes or up across uneven terrain). Turns out, building a legged robot that's more of a benefit than burden isn't so easy.

A robot doggedly moves forward

The 165-pound (75-kilogram) BigDog represents a major step forward for legged locomotion, a problem whose complexity had frustrated engineers, even prompting some to believe it was impossible to solve. How, for example, could a robot know where to place each foot when walking? "The problem seemed too hard; it just didn't seem like it could be done," says Sanjiv Singh, a research professor with Carnegie Mellon University's (C.M.U.) Robotics Institute in Pittsburgh who in 2005 and 2006 worked with BigDog creator Boston Dynamics to develop its computer vision system.

Robot engineers logged some successes in the early 1990s with the Dante 1 and Dante 2 units built to gather gas samples from the Mount Erebus volcano in Antarctica. Both robots were "dynamically stable" because at least three of their four legs touched the ground at all times, reducing the likelihood that the robot would fall, says Singh, who contributed to the Dante 1 mission. Dante 2 operated like two overlapping coffee tables (each with four legs) sliding over each other to slink to its destination—slow and steady but not very nimble.

Boston Dynamics founder Marc Raibert pushed the dynamic stability concept further in subsequent years as he moved from the Robotics Institute to the Massachusetts Institute of Technology (M.I.T.) and then formed Boston Dynamics. (The company declined to be interviewed for this article.) Improvements in a legged robot's agility could only come when the robot could operate each leg independently without sacrificing the machine's stability. Raibert showed this could be done, Singh says, by creating a robot that was able to sense its different body parts, just like an animal, without the use of cameras or laser sensors.

"You know where your body parts are even when you can't see them," Singh says. "When you run, you don't watch your feet the whole time, but you can tell when you're slipping, or when the ground is softer than what you expect. There is a way to encode in robots different gaits that are not based on decision making, you just sort of step [and deal with the consequences]. This is basically the genius of what they've done with BigDog."

The biggest challenge in making BigDog work is "you don't have one joint per leg—you've got four of them," says Robert Mandelbaum, the program manager in DARPA's Information Processing Techniques (IPTO) and Tactical Technology offices who is in charge of the agency's biorobotics program, which includes BigDog. "You've got to navigate a 16-dimensional space and make sure they're all working together to keep its center of gravity." (For more on BigDog, read "Leggy 'BigDog' Robot Set to Step Up for the Military.")

LittleDog's big challenges

What's so difficult about creating autonomous legged robots? In short, "everything," says Tom Wagner, program manager in DARPA's IPTO. Robots such as 4.9-pound (2.2-kilogram) LittleDog are designed to sense the world around them, make decisions based on the information they gather, and then attempt to take some action based on this information. "There are fundamental research challenges that lie in all of these areas, [such as] whether the system can differentiate tall grass from a barbed wire fence, plan its path accordingly, and then follow along that planned path even when the terrain is uneven and difficult," he adds. For an autonomous system like LittleDog, all of the difficulties with perception, cognition and action are combined with the engineering challenges posed by the mechanical system.

Put another way, legged robots must be taught how to walk, and different surfaces require different adjustments. It is a lesson that animals pick up at an early age by using their brains to understand what works and what does not during the learning process. (Walking on carpet is a lot different than trying to navigate a slippery tile floor.) "Look at a gazelle—all of its software is in its brain," says James Kuffner, an associate professor at C.M.U.'s Robotics Institute, one of six teams of robotics researchers (along with the Florida University System's Institute for Human and Machine Cognition, M.I.T., Stanford University, the University of Southern California and the University of Pennsylvania) that DARPA asked to improve on the same basic LittleDog quadruped robot platform, built for them by Boston Dynamics. (For more on LittleDog, read "DARPA Pushes Machine Learning with Legged LittleDog Robot.")

The ultimate robot

A robot's surroundings can prevent it from doing exactly what it is told to do. When a computer uses artificial intelligence to play chess, there is no uncertainty about where the pieces are and where they can be placed. That is not true in a real-world environment, which has endless possibilities that no amount of programming can ever anticipate. To get around this problem, BigDog does not use cameras or laser sensors to determine its location. Instead, it steps first and then reacts to the terrain. This means it must very quickly determine its position at any given time, compare that with its desired position, and immediately take corrective action based on the difference between these two. "BigDog is reacting at 1,000 times per second as it tries to keep its center of gravity," Mandelbaum says. "It only finds out about terrain after the fact."

BigDog does this by sensing the positions of its joints. As it moves, the robot will bend one of its knee joints and then straighten it; if the knee joint fails to straighten, the robot determines that it cannot put weight on that leg without falling over. Using onboard sensors that indicate whether it is tilting left or right or is otherwise unbalanced, BigDog's software checks its weight distribution and relies on its other legs to regain its balance. The strategy seems to have worked: The robot is able to avoid falling when it is on ice and after being kicked in the side.

In addition to controlling BigDog's joints, other major challenges are making the robot durable (so it doesn't break down in the field), efficient (it needs to be able to carry its own fuel and/or batteries in addition to military equipment), and quiet (its two-stroke engine is noticeably loud and may require mufflers).

Gait control—determining when to walk, trot, run, etcetera—will play an important part in BigDog's success, Mandelbaum says. "When a kangaroo achieves maximum speed, it recovers 93 percent of the energy expended," he adds. With that sort of return on energy expenditure, BigDog could get away with having a smaller and possibly quieter engine; its current power plant produces a loud, mind-numbing drone when in operation.

In the end, having robots that can walk like animals means building ones that more closely mimic them, both in the way they move and the way they think. A handful of other robotics researchers—including those at Japan's Kyoto Institute of Technology—have over the past decade been developing quadruped robots, but none appear to have BigDog's high levels of adaptability, balance and perseverance nor LittleDog's intelligence and awareness. In the end, the U.S. military wants robots with all of these traits to accompany its troops on the ground.