Researchers build robot which can learn & evolve

Posted on 13 Aug 2015 by Michael Cruickshank

A team of researchers and engineers has built the first ever robotic system which can evolve its design over time.

A collaborative project between the University of Cambridge and ETH Zurich built a robot which produces successive generations of smaller robots.

This ‘mother’ robot observes and tests small ‘child’ robots built from plastic cubes with internal motors.

Without any human input, aside from a single overarching command to produce robots cable of movement, the system was able to slowly evolve its designs.

Each of these ‘child’ robots have their own unique code which stores their best traits, before being transferred to a new generation of similar, but slightly optimised robots.

In order for the ‘mother’ to determine which robots were the best, each was tested on how far it travelled from its starting position in a given amount of time.

The most successful robots in each generation remained unchanged in the next generation in order to preserve their abilities, while random mutation and crossover were introduced in the less successful children.

Over the course of 10 generations of these robot children, the devices produced were at least twice as capable as those from the first generation, showing a clear improvement beyond pure luck.

“Natural selection is basically reproduction, assessment, reproduction, assessment and so on,” said lead researcher Dr Fumiya Iida of Cambridge’s Department of Engineering in a press statement.

“That’s essentially what this robot is doing – we can actually watch the improvement and diversification of the species.”

The researchers themselves were looking to see if it was possible to mimic natural evolution within robotic systems.

“One of the big questions in biology is how intelligence came about – we’re using robotics to explore this mystery,” said Dr Iida.

While the researchers themselves have not yet hinted at any commercial applications of this technology, it is possible that a similar approach could be used for rapid prototyping and part optimisation.

Through a combination of computer simulation and this pioneering ‘evolutionary’ testing, it could be possible for a robot to manufacture objects with entirely novel designs and superior capabilities to what is currently on offer.