Genetic robots: Golem or how to create evolutionary robots that make themselves

A real first: robots that have acquired the ability to walk after a long evolutionary process have managed to build themselves. A first step that could pave the way for the future advent of robots that can evolve and rebuild themselves according to the complex tasks they will have to perform.

"What I'm trying to do is distort the thing far beyond its normal appearance; but while distorting, I want to get it to attest to that appearance which is its own. "Francis Bacon.

"The truth of life is in the impulsiveness of matter. Man's mind is sick in the midst of concepts. "Artaud

(Photo credit: ©Lipson&Pollack)

Tell him to "build me a robot that can climb stairs" and, using only a few simple and already existing components, the computer, with its "genetic" algorithm, will find how to assemble them, like a Lego game. It would be enough to have an automatic assembly system - which is already possible - to create a population of adaptive robots?

At the Euromold trade fair held annually in Frankfurt, Andreas Fischer and his team showed curious assemblies of tubes a few tens of centimetres in length, held and articulated together by ball-and-socket joints. Multiple configurations are possible and the assembly can be made mobile by means of a bellows with which some tubes are equipped. This is not a new set but a presentation of research carried out at the Fraunhofer Institute for Engineering and Automation (IPA). These robots were not designed by a human brain but by software. The result is a file in the classic CAD format and this computer tool can therefore be interfaced with an automated production system.

Andreas Fisher's team calls these creations "genetic robots". Why genetic? Because their development is due to a "genetic algorithm". The software randomly generates multiple possibilities of assembly, models the result obtained, confronts it with an objective and selects the best formulas. In short, a principle of mutation-selection.

That is what this is all about. The German team, which talks about "bionics", is using software developed by Hod Lipson and Jordan Pollack for their "Golem Project". Hod Lipson has been working for a long time on robots capable of replicating themselves and studying their own physical characteristics (the same has recently produced software capable of deducing physical laws from experimental facts, in short a scientific robot...). The self-replication of machines has been a common field of study for several years.

A robot from the Golem project. Its solution to move on the sand is original but efficient. The Golem Project

Robots that get by

As a starting point, the algorithm starts from a simple and concise specification: for example "to walk as efficiently as possible on a flat surface", explains Andreas Fischer. But it is also possible to ask to cross a staircase or even to swim in the water. The software then simulates physical phenomena, taking into account gravity or friction forces.

This principle of self-teaching using simulators is not new either. The AnimatLab team (now part of Isir), led by Agnès Guillot and Jean-Arcady Meyer, has already achieved many such achievements: virtual ants have learned how to walk (on six legs or less) and other software has learned how to fly a helicopter in hovering flight or how to balance a rod on the tip of a finger (or even two rods). In the end, the algorithm offers several possibilities between which the human user will make the final choice. All that will be left to do is to start manufacturing. "The algorithm can find surprising solutions - "mutations" - that the human designer might not have imagined," says Fisher. That's what the AnimatLab virtual ant did when it was able to move around with just one leg...

(Source : admiroutes.asso.fr)

{Jacuzzi on}