Editorial: “Darwin trumps self-obsession in robotics“
PICTURE what a robot might look like 50 years from now. It’s mowing the lawn, or helping you with the housework. Now, what shape is it? What is it made out of? Does it have arms, legs and a head?
Advertisement
Chances are, what you were imagining does not have a squishy body and tentacles – but such a creature would be closer to the real future of robotics. For many tasks that we actually want robots to do, a hard body or humanoid shape just isn’t cutting it. So researchers are rethinking the fundamentals of what a smart machine is.
Take the robot being built by Cecilia Laschi and her colleagues in the Italian city of Pisa. At the cost of a cool €10 million, they are building a soft, rubbery, and . Its tentacles are made from rubber, mesh and cables, and move in an elegant, natural-looking way underwater. With little brainpower, these arms can grip things that the most advanced robots would struggle to grasp. Researchers are excited that this octopus – and a crop of similarly squishy robots – represents a new era of robotics.
What is really game-changing about this breed is the way in which they are intelligent. They were born thanks to a rethink of how we should design intelligent machines – an approach called “morphological computing”. Its proponents argue that it is not only a robot’s brain that can compute, but its body too. The way a limb, torso, or whisker interacts with its surroundings can be optimised to enhance its computational abilities – and therefore how smart the robot is.
The designers of this new generation are building machines with all sorts of unexpected shapes and odd materials – making them optimally suited to their tasks and environment. If the octopus and its squishy brethren that have emerged so far are anything to go by, the real robots of the future will be softer and smarter than ever – displaying an altogether different kind of intelligence to what we have come to expect.
Until recently, the conventional view of intelligence has been based on the way we tend to think about ourselves – a central, powerful brain commanding and controlling how we breathe, walk, talk and so on. The same approach to robotics initially looked promising. Industrial robots have transformed the way we make everything from cars to computer chips. And in virtually every one of them, each movement of an arm and every twist of a joint is controlled by a central processor.
“The conventional view of robot intelligence – with a central, powerful brain – is fatally flawed”
That works well enough inside settings like factories, a predictable environment where specific actions have to be repeated over and again. But ask such a robot to navigate a maze or tie a pair of laces and it will blow a fuse. As soon as a robot faces a situation it doesn’t recognise, it is paralysed by inadequacy. It is simply impossible to program for every eventuality. So a growing number of roboticists, biologists and engineers are beginning to say that this approach is fatally flawed.
Body smarts
To understand how building robots can be done differently, just think about the way you walk, says , a roboticist at the Artificial Intelligence Laboratory at the University of Zurich in Switzerland.
During this process, the brain doesn’t monitor and control the trajectory of each ankle, knee and hip joint. Instead, it simply changes the stiffness of the leg muscles. The muscles have low stiffness when the leg swings forward and high stiffness on impact with the ground.
Other than that, the brain lets the local dynamics take over. The knee joint simply swings passively, but the design of the knee, the materials from which it is made and the laws of physics all combine to do the rest.
In a sense, the morphology of the body – its shape and substance – perform a kind of computation to control what is going on. The brain simply outsources control to this embodied intelligence. “This relieves our brains from having to deal with all this low-level stuff,” says Pfeifer.
Researchers studying human cognition are now realising how important this body-intelligence is for us – even in shaping our social or mathematical abilities (Â鶹´«Ă˝, 15 October, p 34)
These ideas are changing the way roboticists think about their craft. And the artificial octopus being built in Italy is at the vanguard. “An octopus is a very simple creature and so shouldn’t be very intelligent,” says Laschi, a . “The puzzle is that the behaviour of an octopus is very rich. You could say it is intelligent.”
“An octopus has rich behaviour but it is a very simple creature. It shouldn’t be intelligent”
Laschi talks about these creatures with an obvious affection. She describes their impressive range of behaviour. They grasp objects to eat; they clean their bodies, brushing their skin with their arms; sometimes they hide or collect stones and shells to build a shelter. “When you walk past their tank, they look at you and follow you through the glass. They even try to touch you,” she says. Many other researchers have observed the elaborate, complex behaviours of octopuses and also concluded that these creatures possess a surprising intelligence (Â鶹´«Ă˝, 11 June 2011, p 36)
“But that’s strange, no? An octopus is a mollusc, like a snail,” says Laschi. “How can a creature so closely linked to a snail be so intelligent?”
She and her colleagues came up with a counter-intuitive hypothesis. They think that the octopus is intelligent because of its amazing body, with its eight legs and powerful eyes; that the physical abilities of these organs are what produce this very rich interaction with the environment.
That takes a moment to sink in. After all, conventional thinking is the other way round: an octopus has some central processing capability, some kind of brain that is intelligent, and it demonstrates that intelligence by using its body to interact with the world.
Laschi and her colleagues, including Pfeifer, are turning this notion of central processing on its head. Their idea is that what we think of as intelligence is the body itself, and its capability to interact with its environment. So the octopus’s intelligence comes not from a powerful central brain but is instead embodied in its impressive tentacles and body shape. That’s why snails and octopuses are so different. She points out that if you had 10 snails, chances are you would see similar behaviour in them all. But not with their many-legged cousins. “We had several animals and they all demonstrated different, rich behaviour,” she says. And the difference is their bodies.
To test their hypothesis, Laschi and her colleagues set out to build a robot octopus in a way that captures this embodied intelligence. It would be unlike anything roboticists have handled before.
That task has proven far from simple. Before the team could even get their hands wet, they had to find materials that matched the properties of a real octopus leg, and then model mathematically the way that muscles control it. “The mathematical model was particularly challenging,” says Laschi. “Roboticists are not used to dealing with the infinite degrees of freedom that you get with an octopus leg.”
Eight legs good
With the mathematical model complete, the team built a single leg, which was soft and very flexible. It is a long, thin cone of silicone, supported by an internal wire mesh, and is operated by tugging on cables – some cables shorten the arm when pulled, and others bend and coil the arm when tugged (see diagram). Various combinations of pulling and releasing these wires give the arm all the movement it needs to reproduce real octopus behaviour. When the shortening cable is released, the arm springs back to its extended shape, enabling a pushing action. The coiling cables allow the arm to curl and twist.
Out of the water, the arm is floppy and helpless. But place it in its tank and something extraordinary happens. Its movement suddenly bears an uncanny resemblance to the reaching motion of an octopus. In fact, it almost looks alive.
And that’s the trick. With morphological computing, it’s not just the shape and substance of a body that’s important, it’s also the interaction with its environment that is crucial. It has the dexterity and grip to grab hold of all sorts of different objects placed into its tank. It can also push against the bed in the same way that octopuses use to “walk”. And all with relatively little programming.
The plan is to build an eight-legged robot octopus within two years. But what the team is already learning is invaluable, and has garnered them an invitation to work on a new generation of soft robot surgeons. A team at Kings College London is studying the possibility of a soft arm that would enter the body through the bowel, oesophagus or an incision. It would pass comfortably past soft tissues and organs, then when in place be made to harden, providing some rigidity for whatever surgical procedure is necessary. This could drastically reduce the risk of damage compared with using all-rigid instruments.
Another example in which the principles of morphological computing have been applied can be seen in a robot that can pick up objects in a much smarter way than its predecessors. One problem that roboticists have wrestled with is the challenge of grip: to design a robot that can grasp, say, a delicate glass. A robot with a human-like hand needs to know first how to orient its fingers to apply a decent grip. It also needs to know how hard to grasp the glass (lest it break) without holding it too gently (lest it drop).
That challenge has occupied the minds of engineers without reaching a satisfactory solution. The number of variables, such as the size and shape of different glasses and the finger trajectories needed to grasp them, is too high for the central processing approach.
At Cornell University in Ithaca, New York, Hodd Lipson and colleagues have a different tack. They have created a universal gripper that can pick up more or less anything without knowing a single thing about the object. It uses the same kind of emergent, embodied intelligence seen in Laschi’s octopus.
This robotic gripper is deceptively simple. It consists of an airtight rubber bag filled with grains, in this case ground coffee, connected to a pump. The bag is soft and pliable: place it over an object and the grains form loosely around it. Suck the air out of the bag and the grains jam against each other, forming a rigid mass. The grip comes from three factors. First, if the object and the gripper form an interlocking shape, that holds it firmly in place. There is also helpful friction between the bag and the surface of the object. Finally, there may also be a suction effect if the bag forms an airtight seal around part of the object.
Unlike other robotic grippers, Lipson’s bag can pick up a huge variety of objects – even delicate ones like raw eggs or awkward shapes like a metal spring. The amount of computation that the traditional robotic hand would require to carry out these tasks is huge. Lipson’s bag performs it using only a clever design and the materials from which it is made. It’s essentially just a bag of coffee and a suction pump, but its abilities make it smart.
The morphological computing approach is altering how we think about intelligence and what it is. Consider the following example. Not so long ago, people were amazed by the ability of a mechanical robot to solve a maze. After all, it’s a problem that can give human intelligence a run for its money. But last year, Bartosz Grybowski and colleagues at Northwestern University in Evanston, Illinois, performed this feat of maze-solving in a surprising new way. He and his team created a maze the size of a postage stamp and filled it with an alkaline solution. They then placed a gel containing a strong acid at the exit of the maze. As the acid diffused into the structure it set up a pH gradient.
Blob-nav
Finally, they placed an acidic oil droplet at the entrance to the maze and watched what happened. To their amazement, the blob of oil navigated its way to the exit. What’s extraordinary is the way the blob moved, sometimes “confidently” and at other times “hesitantly”, taking a wrong turn and then retracing its path. To our eyes, it looks intelligent.
But the blob is simply following the pH gradient. It doesn’t rely on intelligence in the traditional centralised sense. The solution to the maze is encoded in its structure and the behaviour of the oil droplet simply emerges. The trick here is in the design: to exploit the materials and the environment so that a solution to the problem emerges.
The oil droplet is no robot, of course, but the principle in which its “intelligence” emerges are the same as those being exploited by the robotocists using morphological computing. In any case, it is time to reconsider the notion that robots are hard-bodied creations with actuators and metallic bodies.
The shapes and materials that the morphological computing approach generates could become increasingly exotic. Before building anything, Pfeifer, Lipson and others routinely allow their designs to “evolve” through successive generations inside computer models. They do this because it’s so difficult to anticipate the emergent behaviour of their potential products before they build them.
The approach is inspired by nature, whereby evolution has optimised arms, legs and tentacles and so on. If an organism’s behaviour is dysfunctional, it doesn’t survive: every time an octopus reaches for an object, it is exploiting the collective “knowledge” of its ancestors.
The upshot of this evolutionary approach will be robots that are ideally suited to the task at hand, but often what no human designer would have come up with – artificial creatures with quite unexpected shapes and materials.
The next generation of artificial creatures will be smarter than ever before, but chances are they will have bodies nothing like ours, and as a result, they will be intelligent in ways that we might struggle to recognise. Eventually, the idea of building humanoid robots in our own image may strike our descendants as rather self-important, if not seriously flawed.
