Is there an insect that can learn

Insects are robots

The modular body structure and the modularized behavior from rigid action patterns make insects interesting for robotics researchers

Insects have mini-brains that, in most cases, have fewer than a million neurons. Insects behave largely "reactive", i.e. there is an immediate and almost automatic coupling between sensory impressions and the triggered action. For this reason, insects are exciting subjects of study for both entomologists and robotics researchers.

I have occasionally debated the subject of this article with my colleagues in biology. You can't accept the statement that insects are supposed to be robots. If you contradict, I appease you with the sentence: "I mean, very good Robots. "I know that with such theses I could move on thin ice and come close to Descartes, who classified insects as unconscious automatons, almost only understood as" clockworks "that mechanically unwind their lives and where they Soul, ie the "thinking substance", is missing.1 However, I think that insects are great creatures, real masterpieces of miniaturization, and that one can gain valuable insights for robotics from their studies.

Take, for example, the little wasp with the scientific name Megaphragma mymaripenne. It is the third smallest known insect: fully grown, it is only a fifth of a millimeter long. The total number of neurons in its nervous system is also astonishing: this wasp species needs 7,400 neurons, 2, while the housefly needs around 300,000 neurons and a bee needs around a million neurons. When you consider how complex a wasp's behavior is, you can marvel not only at the frugality of nature, but also at the efficiency of insect brains. The Megaphragma mymaripenne has the smallest number of neurons of all actively flying living things.

One of the reasons why insects are so successful (they are said to make up almost half of the animal biomass on the mainland and make up 60% of all animal species described) is that they do not weigh up their actions for long. Your nervous systems are clear, as sensors and actuators are connected over short distances. In most cases, their behavior consists of hard-wired "patterns of action", i.e. chains of fixed actions that lead to a desired effect. In robotics, such a direct coupling between sensors and actuators is called a "reactive architecture". Researchers like Rodney Brooks think that even human intelligence can be traced back to a hierarchy of such reactive patterns of action. From a technical point of view, it is therefore easier to use intelligence bottom-up instead of top-down to design. As a result, insects are very insightful for cognitive scientists, since intelligent behavior in these living things is easiest to examine in the laboratory.

Sphexishness

Douglas Hofstadter and other computer scientists already take the wasp as an example Sphex ichneumoneus 3 One of their behavior patterns belongs to the cabinet of curiosities in entomology. The biologist Jean-Henri Fabré, the "homer of insects", has in his Souvenirs entomologiques from 1879 already described bizarre experiments with digger wasps that drag paralyzed grasshoppers to the nest. The grasshopper is placed in front of the nest entrance and the wasp then inspects the nest. If the grasshopper is removed a little from the nest entrance, the wasp comes out, grabs the grasshopper, brings it back close to the entrance and then flies into the nest hole for another inspection.

The perfidious observer can easily remove the locust from the nest each time and the sphex wasp always performs the same ritual without noticing that the nest has already been repeatedly checked. Such an automatic, and one would almost say "headless" behavior, is what Hofstadter calls "sphexishness". Meanwhile, philosophers are discussing such experiments and the opposition between free will and sphexnic Behavior. In order to save the honor of the wasps, however, it should be noted that not all wasps are fooled as often as with Fabré. Others are even a little more awkward. 4

And yet: This is typical for insects, not only the modular body structure, but also the modularized behavior based on rigid patterns of action. Not all sensory impressions reach the brain first - some are connected directly to the corresponding muscles. This also happens in humans, e.g. the trip reflex triggered by neurons in the spinal cord. Before we are aware that we have tripped, our leg has already started the necessary "rescue operation".

Another example of this immediacy of insect behavior can be observed in the so-called phonotaxis of grasshoppers. The males lure the females with their rhythmic "singing". The male sounds can be recorded in the laboratory and played back for the females. The noises can be played through loudspeakers either to the left or right of the female. The female then automatically changes the direction of movement (is attracted by the singing) and you can literally "control" it, no matter how stochastic the sequence is. This movement to the sound (phono-taxis) is an action that is permanently interconnected in the behavior of the grasshopper, but which is only released in the mating context.

And Unleashing is the right word. Many other automatic biological systems are modular. It is characteristic that the modules a) are connected in parallel, b) are data-driven and c) they only report when they are no longer inhibited. Our own visual perception is a good example of this. We recognize the world automatically and unconsciously. With every look we see objects, faces, colors, movement. All modules of the visual recognition are triggered by the availability of the data (the image), all of them "calculate" and generate their results, but the attention only selects from them what interests us at the moment. The corresponding module becomes unleashed and may report its conclusions to the consciousness. We then see something specific5 while many other visual details are suppressed.

It is similar with insects, where many reflexes and behavioral patterns are constantly evaluated and provided. Depending on the context, one of them will be released. In insects that have had their heads severed, for example, this may mean that more neuronal activity can be measured via the axons to the body than in whole insects. Since the brain can no longer inhibit behavioral modules, many reflexes are released. As a programmer, such a system can be understood as parallel running subsystems that perform their calculations at the same time. But only one result is picked from this. The control hierarchy in insects seems to be rather flat compared to our own behavior, which consists of motivations with sub-motivations and further sub-sub-motivations.

How the fly flies

The most astonishing example of how far one can get with permanently wired automatic reflexes is for me the common housefly. Today, when small flying robots are being developed at many universities, none of them can match the flying skills and acrobatics of the housefly. Houseflies are small and accordingly have small brains. You'd be willing to think they're stupid - until you try to catch them.

The eyes of the flies surpass the human visual system not in the number of "pixels", but in the speed of perception. When flying, the so-called "flicker fusion" only occurs at 200 to 300 images per second. In other words: cinemas for flies would have to show feature films at 300 frames per second, while we humans already have the impression of seeing steady movements in films from 24 frames per second. For flies, our films are rather boring slide shows with not much going on. Such high rates of flicker fusion are vital for flying animals, but the housefly comes almost to the limit. It is also clear why: In order to avoid predators or quickly chase other flies through the air for a mating, ultra-fast visual reflexes are necessary. The evolutionary process does the rest, especially in a rapidly reproducing animal like the fly.

It really is the case that the fly is fast because it can also see quickly. The flight muscles must, however, also be controlled quickly and precisely. An engineer would immediately say: "What this fly needs is a fast gyroscope to determine the 3D position of the body at lightning speed." Indeed, the housefly flies around with an extremely sophisticated gyroscope.

House flies are two-winged birds (Diptera), they only have one pair of wings, while many other flying insects have two such pairs. In the course of evolution, the fly's rear wings have degenerated into small swinging bulbs (they are called holders), which also swing during flight (in phase opposition to the wings). However, because the holders are so small, they do not generate any buoyancy: Their function is rather to determine every change in direction of the body (see Fig. 1). The holders swing in a plane relative to the body (like a pendulum). When the fly rotates the body, the holders strive to keep the same plane of rotation; but they are carried away by the body. The force occurring at the base of the stem of the oscillating bulb can be processed neuronally and interpreted as a change in position. With this information, the fly can measure and stabilize its flight movement. The brain is not involved: all information flows directly from the gyroscope (i.e. from the holders) to the flight muscles. When elegantly turning around obstacles, the optimal flapping of the wings is generated by the interaction of the swinging arches with the wings. 6

But that's only half the story. The housefly and many other insects orientate themselves when flying on the countermovement of the environment. This is what is known as the "optical flow" in the visual system. A fly that flies against a wall, for example, sees the details of the wall getting bigger and bigger. The world "bursts" in all directions. When this happens, the fly turns its body and stretches its legs to land. Should the fly fly in a tunnel with structure along the walls and the optical flow would be faster in the right eye than in the left, then this means that the right eye is closer to the wall than the left eye. The fly corrects the direction, balances the optical flows in both eyes and can thus fly centered in the tunnel. Even when the head is rolled, an optical flow is created downwards in one eye and upwards in the other. The fly would then know that the body has turned and could counter-steer in order to keep it horizontal again. Today we know, for example, that the "optical flow" is the bees' odometer. When bees fly to feeding grounds, they measure the flown distance by integrating the perceived optical flow of the ground.7 They can then communicate this distance to other bees through the bee dance.

But what does an engineer do when suddenly you have two different "sensors" to measure changes in flight attitude? On the one hand, the fly has the holders, but on the other hand it also has an ultra-fast visual system with which the optical flow of the environment and thus the self-movement of the fly can be determined very precisely (as an engineer would say: "yaw, pitch and roll" the fly can can be determined). Evolution is blind and opportunistic. A neural connection between the holder and the flight muscles is already in place. It is a relic of the ancestors of the fly, in which two pairs of wings were used in a coordinated manner. You can therefore redirect the measured optical flow in the eyes directly to the holders, connect the optical flow measurement to the fly's gyroscope and send a single control signal to the flight muscles.

Instead of building a complicated decision-making mechanism in the brain, evolution simply supplemented the existing interconnection (tweaking would be the appropriate English word). The signal from the holders is thus modulated by the optical flow. Every engineer at Daimler would have to take his hat off for a similar construction with the inertial systems used in cars. But for houseflies it is an ingenious solution. Fig. 2 shows the neural interconnection from the eyes to the holders and from these to the flight muscles as revealed by Dickinson.8 The brain is not very involved - just to provide the optical flow.

"If you think it's too late anyway"

The headline was Gerd Müller's answer when he was once asked what he was thinking about while scoring. That sums up the situation very well if something needs to be controlled in real time. When long chains of deliberations are put in front of them, there is simply no time to react. The flight reflex in cockroaches is therefore triggered in less than 40 milliseconds. Tiny hairs on the rear part perceive the smallest vibrations in the air and immediately trigger the flight reflex. To underline the urgency of the situation even more drastically, the associated neural signal is transmitted to the body via a very thick and fast axon. It's like a taut spring just waiting to be released.

House flies also act like a goalscorer: they react first and then, if at all, think about it (whereby a fly only takes 50 to 100 ms to flee, but a footballer needs at least 250 ms to react). One of the nastiest experiments you can do with flies has been described by Dickinson. If you take a stick with a little glue, touch the back of the fly and lift it up with it, the legs lose contact with the ground. The fly, which is held with the chopstick, immediately begins to flap its wings. This is the fly's fly reflex. It is necessary because, until the entomologists emerged, houseflies without ground contact could not do anything other than fly. And they do this all the time. You don't fly for minutes, but for hours. If they are regularly fed sugar, they will fly on tirelessly for days until their wings tear. It's the right reflex, but not in an insidious context.

Legs that learn

Insects have such a modular structure that you can even replace body parts with prostheses without them noticing much of it. With desert ants you can shorten the legs with scissors and then glue on small wooden sticks. The ants run as usual, only that they may now take longer strides and they make a mistake in navigating because their odometry is mixed up. In ants whose head has been severed from the body by a wasp, the head survives for a few minutes and continues to move the antennae. It is almost as if the components of the insect's body were Lego building blocks that you just need to connect together.

Another experiment with insects that only requires scissors and glue is this: Moths use their antennae like gyroscopes. They swing the antennae in sync with the wings, just as houseflies do with their holders. If you cut off the antennae of a moth, it can no longer fly properly and falls to the ground. But if you glue the antennas back to the antenna stub with superglue, the moth can fly again! The gyroscope is complete again and since the measurement takes place at the point of contact of the antenna with the body (and this point remained untouched), everything is fine again

The most vivid examples of insect modularity are studies like that by Holk Cruse in Bielefeld on the walking behavior of stick insects. Cruse and other scientists have studied the neural circuitry of walking in these insects for decades and formulated general principles

It turned out that the legs of the insects work pretty independently. Each leg, with its muscles and sensors, is something like a small component, six of which are mounted on the insects. The control is modular, parallel and distributed. The coordination of the legs occurs through the interaction of the reflexes (which are connected via inhibiting and exciting neurons) and the connection to the body, with which the legs behave like coupled oscillators. Just like the mechanical pendulum clocks on the wooden floor in a watchmaker's workshop, the legs can optimally synchronize both mechanically and via neural signals.

The individual leg controllers can be modeled with only six rules that describe their interaction: For example, a leg in the air inhibits the lifting movement of the neighboring legs on the same side, so that the insect remains stable. On the other hand, when one leg hits the ground, it excites the neighboring legs so that they are lifted. A leg moving forward is placed in roughly the same spot that the front leg was to meet bumps on the floor. If one leg carries too much weight, the other legs will have longer contact time with the ground in order to better distribute the weight.With his Walknet controller, Cruse has shown how such simple interaction rules, together with local leg control using leg sensors, can reproduce the various step patterns of some insects.

Such models and experiments make the legs of the insects appear almost like creatures in themselves. For example, experiments with cockroaches in which the head is missing are astonishing. The leg reflexes all still work and the individual legs can even learn! For this, a leg is "punished" with a small electric shock if it touches a plate. The leg "learns" to hold a new position in order to avoid punishment. All learning is made possible by a single neuron and Horridge paradigm called 11

Frank Pasemann from the University of Osnabrück has investigated such self-organizing systems in robotics. With neural coupled oscillators he can build robots in which the modules automatically interconnect and generate non-trivial behavior. His student Manfred Hild (from HU Berlin) played it through with humanoid robots and developed humanoid legs that autonomously learn to "do the right thing", e.g. to stretch or to run. So if someone soon sees a single humanoid leg jumping down the aisle autonomously, Manfred Hild must be notified.

These last examples from mechatronics illustrate very well the idea of ​​"subsumption architectures", which are based on the modularity of insects. A robot is built step by step: first, the simple reflexes are implemented. Building on that, the next level is developed, and so on, until you have a complete and useful system. A walking robot, for example, with optimal leg coordination.

Reactive architectures

It is no coincidence that such reactive architectures are used very successfully in the RoboCup competition (the annual robotics soccer championship). Football itself is very reactive: the goalkeeper usually only has to throw himself down at the right moment, the striker has to put his head or leg in the way, everyone just has to move relative to the ball. Even passes deep into the room can be handed in without looking up: you assume that the other player is in the right place.

Perhaps I can win my friends in biology over to the exciting connection between insects and robots by pointing out that today's automobiles are increasingly using the reactive principles of insect modularity. It starts with the vehicle dynamics control with the ESP (Electronic Stability Program). If a vehicle is in danger of swerving, individual wheels are optimally braked by the electronics. Without ESP, an A-Class Mercedes cannot pass the infamous slalom in the moose test. New Volvo vehicles brake automatically at low speeds when an obstacle is detected by the cameras in the car. With VW's lane keeping systems, the car automatically moves into the middle of the lane. With all of these driver assistance systems there is little logic, just simple rules. You don't want to use complex operating systems that then reboot at the wrong moment. Simplicity is vital here.

Even human cognition seems to be divided into two subsystems, one for automatic control and the other for reflected decisions. The first "thinks" quickly, the second slowly. This is what Daniel Kahneman calls system one and system two.12 System one regulates everything automatically, quickly and sometimes with a fixed behavior pattern. System two is always thinking. A tennis player only needs system one for a return. He doesn't have time to think. System One is the Sphex in us.

That's why I tell my colleagues in biology: insects are robots, but a fly would certainly be a wonderful driver and could win Formula 1 straight away. She already has all the necessary cognitive abilities of system one. (Raúl Rojas)

Read comments (63 posts) https://heise.de/-3395234Report an errorDrucken