When the Honda Corporation recently unveiled the latest version of Asimo, a four-foot-tall humanoid playfully named after the science-fiction writer Isaac Asimov, a digital video of the robot generated the kind of enthusiasm on the Internet that is usually reserved for Paris Hilton’s home movies. The new incarnation of Asimo not only walked with a convincing bipedal gait but also accelerated smoothly from a leisurely stroll to a full-on trot. It was mesmerizing—and comic—to see the sleek white machine jogging across a stage as if it were a harried commuter trying to catch a bus.
A jogging robot captures our imagination because we’re easily impressed by skills that mimic our own. But a robot that runs isn’t necessarily better than one that doesn’t. The future of robotics lies beyond mimicking humans and in machines that transform themselves into configurations based on changing circumstances. Some of these machines may resemble creatures from the natural world, but others may be original, concocted to repair a sudden failing or find a way around an unexpected obstruction. You can be sure that the robots that eventually colonize the galaxies or explore the uncharted depths of Earth’s oceans won’t look like a jogging butler. In fact, they won’t look like any single thing at all because their primary talent will be shape-shifting.
Self-assembly and self-repair are defining attributes of complex life. Think of the army of cellular agents, including white blood cells and platelets, that jump into action over a mere paper cut—rebuilding the tissue, warding off infection, and alerting the rest of the body to the wound through the A-delta fibers of the nervous system, which are involved in the transmission of acute pain sensations. DNA has an elaborate system for minimizing errors when it makes copies of itself. Otherwise, multicellular life would be filled with an intolerably high number of defects. And thanks to the encodings of DNA, cells are capable of complex forms of self-assembly, depending on the task that the body requires in each stage of development. The same genetic strand can be used to build a neuron or a white blood cell or a sliver of muscle tissue.
Anyone who has ever struggled to fix a paper jam in a copier knows that most machines aren’t very adaptable. When machines break, they don’t release a host of component parts to heal themselves. They remain broken until someone calls tech support. Likewise, with the exception of the virtual machines of software, most technology isn’t capable of adaptive self-assembly. The fax machine can’t morph into a toaster when you’re in the mood for jam and bread. But a new generation of experimental robots are capable of precisely this kind of self-maintenance and transformation.
The M-TRAN II robot, developed by the Japanese Distributed Systems Design Research Group, looks at first glance as if it’s assembled out of those cheap plastic adapters you buy to plug two appliances into a single socket. The robot’s designers call these white units modules, and M-TRAN—shorthand for “modular transformer”—is made up of about a dozen of them. Each module contains two 2 1/3-inch blocks linked to each other. Each block can rotate 180 degrees around the link that connects it to its mate, and each module contains a magnet that can be switched on and off, enabling it to connect to other modules in the system.
What M-TRAN lacks in animatronic magic, it makes up for with flexible design. The modules can rearrange themselves into countless different shapes and create dramatically different patterns of movement. M-TRAN can configure itself to look like those relentless Imperial Walker transport vehicles from the Star Wars films, marching steadily on four legs. But it can just as readily shape-shift into a long string of modules, allowing it to inch along like a caterpillar or slither across the floor like a snake. Alternatively, it can pull itself into a wheel and roll or creep along the ground with its legs splayed out like a spider’s.
There are two basic underlying designs for modular robots: lattice-based systems and chain-based systems. Lattice systems involve units that shuffle around a 2-D or 3-D grid, like a group of Lego blocks that have come to life. Chain systems resemble robotic arms that can link together, attaching and reattaching themselves depending on the needs of the situation. “Lattice systems are good for self-reconfiguration but not for robotic motion,” Satoshi Murata, M-TRAN’s original designer, says. “On the other hand, chain systems are good for robotic motion generation, but self-reconfiguration is difficult.” That’s because lattices provide fixed coordinates, which limit mobility but are ideal for aligning modules when a new shape is being formed. By contrast, arms in a chain system can move and grab and push; they’re good for dynamic activities, but it’s difficult to coordinate their positions precisely.
M-TRAN is a hybrid of the lattice and the chain approaches. The blocks in each module move around the blocks in other modules following a lattice system, while the joints in each module allow them to flex and link together in chainlike arrays. The hybrid model gives the robot an extraordinary range of potential shapes and movements. The combinatorial possibilities are so immense that many of M-TRAN’s patterns of motion aren’t designed directly by human programmers. M-TRAN’s creators instead use genetic algorithms that allow the robot to discover new ways of moving on its own. The M-TRAN computer cycles through thousands of possible patterns of motion, selecting the most promising candidates, sampling their effectiveness, and then choosing the most promising as the starting point for the next round of candidates. After a few cycles of this artificial natural selection, the software evolves a new pattern that the robot can adopt.
M-TRAN has an even more fundamental kinship with computers. What made the first digital computers revolutionary was not just that they could calculate missile trajectories or factor pi at superhuman speeds but also that they could perform an infinite variety of tasks. We take this entirely for granted now. Your personal computer might be inept at handling photographs, so you install a piece of software, and suddenly the machine has a new talent. Robots and other mechanical devices have traditionally lacked that open-endedness. They’re limited at birth to a specific range of skills, and those skills are notoriously fragile. If a gear fails or a piece of paper jams, the skill disappears.
Because of its capacity to respond to changing circumstances, M-TRAN suggests a way around a limited skill set. “For a single clearly specified task, a purpose-built specialized robot will probably do the best and be the cheapest to produce,” says researcher Craig Eldershaw of the Palo Alto Research Center. “But the big win from self-assembly is adapting to changing demands and environments—in particular, addressing situations and tasks that the robot designer doesn’t even know about at the time of construction.”
Murata sees future uses for M-TRAN’s descendants as space rovers or deep-sea probes, as well as fearless explorers closer to home. “These machines,” he explains, “are ideal for searching tasks in unknown or complicated environments—say, looking for people under debris after an earthquake or fixing leaky valves in polluted areas, such as nuclear plants.” A self-assembling robot might default to a standard configuration—the four-legged walker, perhaps—when embarking on a rescue mission, then shift into caterpillar or sidewinder mode if confronted with a stretch of debris that makes upright walking impossible.
Having a miniature version of natural selection on hand creates new possibilities for self-repair as well. Imagine a version of M-TRAN designed as a Mars rover that suffers damage to three critical modules during a rough landing. Losing those modules might well prevent the robot from executing the standard four-legged walk, but the robot could evolve a new walking strategy with the remaining modules. It might not be quite as elegant as the original strategy, but it’s vastly better than a rover lying helpless on the Martian soil. NASA apparently agrees: They’ve just hired Eldershaw and his team to develop a new modular robot for assembling future space stations.
No doubt we will continue to have robots designed specifically to build cars on automated assembly lines, erect buildings with minimal human assistance (see “The Whole-House Machine,” page 60), or even jog across the front lawn to pick up ournewspaper. But sometimes you want a tool that can do a million things tolerably well instead of one thing perfectly—particularly if that tool knows how to fix itself when the paper jams.