Using evolution to produce well-adapted robots

How does the interaction between body, brain and environment lead to intelligent behaviour?

Guszti Eiben Professor of Artificial Intelligence Guszti Eiben’s work focusses on robots that reproduce and evolve. “Our project has three objectives”, he says. “We, together with biologists and philosophers, are asking fundamental questions about the emergence of intelligence. We are also developing a revolutionary way to design innovative robots for use in factories or mines, for example. And we are laying the foundation for a new form of science communication: a cuddly ‘robot zoo’ that will inspire a wide public interest in evolution and robots.”

Intelligent beings
In nature, evolution has proven that it can produce well-adapted species and intelligent beings. Eiben therefore expects that, if he combines evolutionary theory with computer and robot technology, well-functioning robots will emerge. “We ask ourselves the same questions that biologists ask”, Eiben explains. “Including: how does the interaction between body, brain and environment lead to intelligent behaviour?”

Body shapes
The ‘body’ of the robot consists of the rump, the sensors and the metal and plastic parts making up the gripper arm or arms; the ‘brain’ is formed by the algorithms that control these parts. The research group is using actual robots, virtual robots and combinations of both to investigate what well-functioning robots in specific environments could look like. 

Robot children
Robots must be able to ‘have children’ to be able to evolve. In 2016, Guszti Eiben’s team was the first in the world to succeed in getting a new robot from two four-limbed, crawling laboratory robots. These ‘parent robots’ managed to reach a red light and make wireless contact with one another there. As a result, their algorithms for the control and generation of parts were sent to a program, which mixed them randomly, as is the case in natural reproduction. With the aid of a 3D-printer, the offspring thus received a mixture of parts and control algorithms from the parents.

Guszti Eiben: “In virtual robots, we have already seen that in a flat, desert-like environment, a convenient body shape with the right parts correctly assembled is more important for survival and reproduction than the control of the parts by software. We have also observed that, in flat environments like this, the types of virtual body shapes that emerge are similar to those that occur in nature, i.e. the penguin and salamander shapes.”

Nuclear power stations
The next step is to have (virtual) robots evolve on (simulated) seabeds, in factory rooms or in blood streams, where they have to fulfil specific tasks. What types of body forms and brains will they need to have? The research group is currently engaged in various activities, including cooperation with three English universities on robots that can clean up abandoned nuclear power stations. The robots have to be able to fulfil this task between trenches and in caverns, potentially hazardous liquids and parts of the premises that may have collapsed.

The most important and essential question as far as Eiben is concerned is how evolution, and particularly artificial evolution, works. At the same time, he wants his research to be of practical use: “But if that usefulness only becomes evident in five or ten years’ time, that´s fine too.”