robots

Dario Floreano and Laurent Keller describe experiments that combine genetic algorithms and robots. It's a review essay rather than a description of new research, but unlike most descriptions of "evolutionary robotics", it's actually directed toward biologists instead of AI researchers.

In this essay we will examine key experiments that illustrate how, for example, robots whose genes are translated into simple neural networks can evolve the ability to navigate, escape predators, coadapt brains and body morphologies, and cooperate. We present mostly—but not only—experimental results performed in our laboratory, which satisfy the following criteria. First, the experiments were at least partly carried out with real robots, allowing us to present a video showing the behaviours of the evolved robots. Second, the robot's neural networks had a simple architecture with no synaptic plasticity, no ontogenetic development, and no detailed modelling of ion channels and spike transmission. Third, the genomes were directly mapped into the neural network (i.e., no gene-to-gene interaction, time-dependent dynamics, or ontogenetic plasticity). By limiting our analysis to these studies we are able to highlight the strength of the process of Darwinian selection in comparable simple systems exposed to different environmental conditions.

Some of the simplest machine learning experiments are basically like those used in behavioral psychology -- put the robots in a maze, make them remember where the food is, that sort of thing. Robots are simpler than rats, so the researchers can reverse-engineer the "evolved" software at the end of a series of experiments to see what worked and why:

Interestingly, the driving speed of the best-evolved robots was approximately half of the maximum possible speed and did not increase even when the evolutionary experiments were continued for another 100 generations. Additional experiments where the speed was artificially increased revealed that fast-moving robots had high rates of collisions because the 300-ms refresh rate of the sensors did not allow them to detect walls sufficiently in advance at high speed. Thus, the robots evolved to move at intermediate speeds because of their limited neural and sensory abilities.

Figuring out that particular optimization would drive a team of human programmers crazy. Can you imagine? "Why do they keep running into that wall?!"

On the other hand, dumb selection took a lot of generations to get to that point. You can't say selection was more efficient. If you had a crew of programming grunts and forced them to sit in a room for 100 robot generations, they'd come up with something.

It's quite possible that a human would have come up with much better software, by pushing the robots past the limits of mutations on their "genomes". Selection has its own "sensor limitations", it can get stuck in a local optimum, and depends on the mutation structure to explore the landscape.

It helps if the landscape has some strong correlation structure. That's what came to my mind as I read their account of experiments to make robots cooperate:

However, when the arena contained both large and small tokens, the behaviour of robots was influenced by the group kin structure. In groups of unrelated robots (i.e., robots whose genomes where not more similar within than between groups), robots invariably specialised in pushing the small objects, which was the most efficient strategy to maximise their own individual fitness them (i.e., large tokens provided an equal direct payoff as a small token but were more difficult to successfully push). By contrast, the presence of related robots within groups allowed the evolution of altruism. When groups were formed of “clonal” robots all having the same genome, individuals primarily pushed the large tokens even though it was costly, in terms of individual fitness, for the robots pushing (Video S6).

If you wonder how robots have "kin", it's that they share similar (or the same) genomes. The simplicity of the behaviors suggests a functional explanation for kin selection -- for many kinds of tasks, it may simply be easier to cooperate with other individuals who "work" the same way. Different approaches to the same task may clash.

They describe a similar result for cooperation by information sharing:

Similar results were obtained in experiments where groups of light-emitting, foraging robots could communicate the position of a food source at a cost to themselves because of the resulting increased competition near food. In these experiments, robots again readily evolved costly communication when they were genetically related, but altruistic communication never evolved in groups of unrelated robots when selection operated at the individual level [38],[39].

The next logical step for this kind of research is nano-scale: evolutionary robotics on molecular machines. Which is scary. I hope they have the sense not to train them up by eating biological systems...

There's this old course on the books here, "Human aspects of robotics". I suppose it was taught back in the 80's when robots looked like they would replace all the manufacturing workers. I've often thought that someday it may be revived as with robots as the heroes instead of the villains.

References:

Floreano D, Keller L. 2010. Evolution of adaptive behavior in robots by means of Darwinian selection. PLoS Biol 8:e1000292. doi:10.1371/journal.pbio.1000292

An interesting story from Popular Mechanics about progress in cybernetics, titled "Mind control stories." It starts with the macaque controlling a robot arm by brain implants, and then considers the future:

For Miguel Nicolelis, a professor of neuroscience at Duke University Medical Center, the backbone of mind-machine interfaces is the ability to analyze neural activity. Sure, the system demonstrated at Pitt in May accessed information from 100 neurons at once. But Nicolelis’s lab has managed five times that amount, with data coming from up to 10 different brain structures.

For me, this is the most interesting part:

The main purpose of the walking robot experiment was to demonstrate just how precisely brain activity could be translated, but it produced another interesting result: It actually took less time for the brain signal to travel from the monkey in North Carolina to the robot in Japan than it took to go from the primate’s brain to its own muscles. At any given moment, then, the bot was receiving the command to walk before the monkey’s body did.

I've been reading Ray Kurzweil's book, and it has always seemed to me that a fundamental barrier to the development of effective neural implants is bandwidth: Human brains have evolved to use inputs and outputs at the speed of language, not the speed of electronics. So this idea of accelerating real-world responses and feedback by wiring may suggest substantial plasticity with respect to bandwidth.

I think I'll lecture on this topic in my "Biology of Mind" course this fall.

So, a monkey in North Carolina was controlling a robot in Japan, using only its brain waves.

"It's walking!" Dr. Nicolelis said. "That's one small step for a robot and one giant leap for a primate."

Well, what else did you expect him to say? Maybe "Mwa-ha-HA-HA!"?

Anyway, the study looks kind of cool -- they had a monkey on a treadmill for an hour, got the electrode reading neurons related to walking, and had the monkey watching the robot's legs on a television screen. Once the monkey got used to the idea of controlling the robot's legs, they stopped the treadmill. At this point, even though the monkey had stopped, its brain kept the robot walking.

The next step: virtual robot monkey reality:

In the near future, Idoya and other bipedal monkeys will be getting more feedback from CB in the form of microstimulation to neurons that specialize in the sense of touch related to the legs and feet. When CB's feet touch the ground, sensors will detect pressure and calculate balance. When that information goes directly into the monkeys' brains, Dr. Nicolelis said, they will have the strong impression that they can feel CB's feet hitting the ground.

At that point, the monkeys will be asked to make CB walk across a room by using just their thoughts.

Unfortunately, they will have to move offshore to have virtual robot monkey knife fights.

Product design guru Donald Norman looks at this year's crop of "smart" machines in this NY Times article, and reminds us why future robot sex ain't all it's hacked up to be:

Until recently, Dr. Norman believed in the favorite tool of couples therapists: better dialogue. But he has concluded that dialogue isn't the answer, because we're too different from the machines.

You can't explain to your car's navigation system why you dislike its short, efficient route because the scenery is ugly. Your refrigerator may soon know exactly what food it contains, what you've already eaten today and what your calorie limit is, but it won't be capable of an intelligent dialogue about your need for that piece of cheesecake.

This is like the Woody Allen version of robot relationships. Plus, it's hard to set a mood when the robot controls the lighting:

As he watched our window shades mysteriously lowering themselves, having detected some change in cloud cover that eluded us, Dr. Norman recalled the fight that he and his colleagues at Northwestern waged against the computerized shades that kept letting sunlight glare on their computer screens.

"It took us a year and a half to get the administration to let us control the shades in our own offices," he said. "Badly designed so-called intelligent technology makes us feel out of control, helpless. No wonder we hate it."

I have exactly the same problem with a motion-sensing light control in my office. I have to do some kind of Morris dance around the room to get the light to stay on for more than 10 minutes!

Just wait until the robot gets the TV remote.

I've been telling people this week that there is some sense to which the evolutionary future will be determined by the cultural impact of technological changes -- genetic engineering being the most prominent example.

Now comes this:

[T]here will soon come a day when people fall in love with robots and want them for companions, friends, love objects and possibly even partners for sex and marriage.

That day is imminent, [writer David] Levy writes, especially the sex part. By the middle of this century, he predicts, "love with robots will be as normal as love with other humans, while the number of sexual acts and lovemaking positions commonly practiced between humans will be extended, as robots teach more than is in all of the world’s published sex manuals combined."

Well, that's one more thing, isn't it? If you're more likely to fall in love with a robot, will you be less likely to have children? And if so, will that mean that over many generations, robot-revulsion genes will be selected?

I'll tell you what, if they make Haley Joel Osment-looking robot children, I'm already revulsed!

Neurophilosophy has really come to life in the last few weeks. A post earlier this week described the neural circuitry that controls swimming in zebrafish, from work published in Nature. Today's post takes the evolution of motion up to tetrapods, with a description of a robotic salamander and what it tells scientists about motor control systems.

And this post about rat metacognition covers the Current Biology paper by Foote and Crystal so I don't have to:

Jonathan Crystal and Allison Foote, of the University of Georgia’s Department of Psychology, taught rats to associate two different auditory stimuli with different levers. A short burst of static, lasting around 2 seconds, was associated with one lever, and a longer burst, lasting up to 8 seconds, with another. In the second phase of the trials, the sounds were played back to the rats. When the lever associated with each sound was correctly pressed, the rats were given a large reward - 6 food pellets. But if the wrong lever was pressed, they received no reward. The rats were also given the option to decline taking the test - they learnt that they could retrieve a smaller reward - 3 food pellets - without making a decision about which lever to press, by poking their snout through an aperture in a food trough.

During the test phase, the rats were presented with the short and long bursts of static, as well as with bursts of intermediate length, and their responses were recorded. When the length of the sound burst was unambiguous (i.e. either short or long) they ignored the food trough and pressed the lever associated with the sound, so that they received the large reward. But when sounds of an intermediate length (approximately 3 seconds) were played, the rats frequently declined to take the test, and chose instead to retrieve food pellets from the food trough, suggesting that the rats knew that they did not know how to respond in the duration discrimination test.

The paper concludes that the rats have a concept of what they know they know -- that is, a metacognitive concept. My students this week told me that rats are smart; I suppose it's true enough.

References:

Foote, AL, Crystal, JD. 2007. Metacognition in the rat. Curr. Biol. doi:10.1016/j.cub.2007.01.061

Signs that the Japanese robot industry has gone too far:

Researchers at NEC System technologies and Mie University have designed the cute little guy to the right: a metal man gastronomist, "an electromechanical sommelier", capable of identifying wines, cheeses, meats and hors d'oeuvres. Upon being given a sample, he will speak up in a childlike voice and identify what he has just been fed. The idea is that wineries can tell if a wine is authentic without even opening the bottle, amongst other more obscure uses...like "tell me what this strange grayish lump at the back of my freezer is/was."

But when some smart aleck reporter placed his hand in the robot's omnivorous clanking jaw, he was identified as bacon. A cameraman then tried and was identified as prosciutto.

One of those things that says, "Please stop reading now" :

Doctors currently explore the gut using endoscopes, which have to be fed through the body, or "camera pills" that must be swallowed by a patient.

A pill capable of wriggling through the gut on its own could be a valuable tool, says Andrew Gardner, an independent medical imaging expert at University College London.

Yes, you have to go over to New Scientist to see the picture. It's two tail spines short of a centipede. And it's crawling through pig guts right now. Eeeeww!

"If something this complicated goes wrong, it could be very hard to get out."

Eeeeww!

It seems clear that we have only one hope against superintelligent fearless killer mice: Robots that carry out our telepathic commands!

Happily, Honda Motor Co. is preparing for the day:

In a video demonstration in Tokyo, brain signals detected by a magnetic resonance imaging scanner were relayed to a robotic hand. A person in the MRI machine made a fist, spread his fingers and then made a V-sign. Several seconds later, a robotic hand mimicked the movements.

I don't know about robotic hands pre-equipped to make a victory sign...

What Honda calls a "brain-machine interface" is an improvement over past approaches, such as those that required surgery to connect wires. Other methods still had to train people in ways to send brain signals or weren't very accurate in reading the signals, Kamitani said.

This is interesting in that computers are being used to capture the signs in brains that are not consciously produced, as opposed to methods of training people to make brain wave patterns that can be externally perceived with much cheaper and lighter equipment. It presupposes a future in which methods of scanning brains become much cheaper and ubiquitous. I suppose they might if there were reasons -- and maybe a multipurpose wear-anywhere brain scanner really would sell if they had lots of uses.

Like for voting on "American Idol"...

The BBC is running this article about a new study that evaluates the bipedality of A. afarensis using robotic design software:

Now, a team of scientists from around the UK have used computer robotic techniques to work out the most energy efficient gait for afarensis based on Lucy's skeleton and the Laetoli footprint trails.

They claim to have cleared up the debate by finding that, based on their model, Lucy almost certainly did walk tall.

There has been a long-standing debate about how human Lucy was
"Assuming that the early human relative Australopithecus afarensis was the maker of the Laetoli footprint trails, our study suggests that by 3.5 million years ago at least some of our early relatives - despite their small stature - could sustain efficient bipedal walking at absolute speeds within the range shown by modern humans," co-author Weijie Wang, from Dundee University, told the Scotsman newspaper.

So what we seem to have here is a computer software equivalent of early 1980's science. Perhaps they programmed Owen Lovejoy?

The paper (abstract) is a little more interesting than the BBC description. It does a kind of optimization modeling to find the speed and style of locomotion with the lowest energy cost. That lowest-cost speed was associated with a human-like gait at around 1 meter per second (m/s). Here is the logic:

Rather than trying to interpret the behaviour of such species by a combination of analogies to humans in certain anatomical regions with analogies to other apes elsewhere, it seems sensible to adopt a reverse-engineering approach and determine what kind of locomotion a particular set of body proportions were best 'designed' to perform. Since the locomotor system is concerned primarily with the application of external force by the body, simulation techniques drawn from mechanical engineering are a potential means of predicting the significance of differences in proportions for the motion and force characteristics of bipedal locomotion (Sellers et al. 2005: 2).

The authors relate their gait findings to the preserved Laetoli footprint trails, and find something very interesting. The trails have footprints that are very close together -- especially for the larger, possibly dual G2 trail. This has previously been interpreted as meaning that the walking speed of the larger individual who made this trail was very slow (Alexander 1984; Charteris et al. 1984) -- only around 0.7 - 0.8 m/s. The model in this study predicts a faster speed of around 1 m/s, which would be close to the optimal walking speed estimated for Lucy's proportions.

They do not address a relevant question, which is whether the larger trail was made by an individual larger than Lucy, who might have had a different optimal walking speed. Indeed, there is a basic assumption of monomorphism. It is partly covered by the fact that living people don't exceed 1.0 m/s for their average walking speed by very much, so the variation between large and small A. afarensis individuals may have been slight. The basic finding seems to be that shorter legs have an optimum gait that involves more, slightly shorter, steps, while longer legs take fewer steps more slowly. This isn't surprising based on a pendular model: shorter pendula have shorter periods, and longer steps with a shorter leg must require more energy-wasting up-and-down motion.

An important weakness of the model is that it considers costs only due to motion in two dimensions: forward and up-and-down. The wide pelvis of A. afarensis might be expected to exert greater costs in a side-to-side dimension compared to recent humans, and that energy effect is not considered.

But the bottom line:

Thus, within the limits of our model, and assuming that Taylor and Rowntree's (1973) data are reliable, the bipedal performance of Australopithecus afarensis, as predicted by our model is not only much closer to that of modern humans than to that of bipedally walking great apes, but at normal walking speeds, shows a clear speed/cost advantage over chimpanzee quadrupedalism. Climbing remains significantly more energetically expensive than terrestrial quadrupedalism for chimpanzees, despite their musculoskeletal adaptations (Pontzer and Wrangham 2004). Pontzer and Wrangham (2004) have shown that the costs of locomotion in chimpanzees are nevertheless dominated by terrestrial walking, because of very high daily travel distances, versus only limited use of climbing. If we make the major (and quite likely incorrect) assumption that the African apes existing at the time of A. afarensis were ecologically, morphologically and physiologically similar to modern common chimpanzees, then our data would tend to support Rodman and McHenry's (1980) argument that the adoption of bipedalism offered energetic advantages to early human ancestors (ibid., 9).

In any event, it won't quiet doubters:

However, Professor [Christopher] Stringer believes the controversy will not vanish overnight.

"There are still some people who argue that, looking at the anatomy of the foot bones of afarensis, that they were unlikely to have made the Laetoli footprints," he told the BBC News website.

"So it doesn't end the argument because there is still the possibility that there were different creatures around at the time."

No doubt soon, Kent State will have a droid afarensis army to finally crush this dissent and bring order to the galaxy.

References:

Sellers WI, Cain GM, Wang W, and Crompton RH. 2005. Stride lengths, speed and energy costs in walking of Australopithecus afarensis: using evolutionary robotics to predict locomotion of early human ancestors. Roy Soc Interface Online