Mother Robot generates offspring emulating natural evolution

Mother Robot

 Editorial Staff

 27 August 2015

“Mother Robot” generates offspring emulating the natural evolution processes

Researchers from University of Cambridge in collaboration with ETH Zurich have developed a robot able to assemble its children, test them, and use results to improve performances of the next generation.

Evolution is the natural phenomenon through which heritable characteristics in biological systems change over generations. This process grants species adaptation and survival in their environment over time.

On the contrary, today’s machines are strictly bound to their initial shape and configuration.

Several attempts to reproduce natural evolution mechanisms in robotics have already been made, but most of them have used computer simulation, where optimization relies on the precision of physics models.

Once the process is transferred from virtual to real, these methods always suffer the so-called “reality gap” due to the limited accuracy of physics models when tested in the real world.

University of Cambridge experimented a new method, which allows for physical adaptation and which is entirely executed in the real world from design, through optimization, to iterations.

The experiment resulted in confirming the feasibility of a model-free evolution process of a physical entity able to generate modifications and adaptations in order to achieve species improvement.

The research approach

Researchers of the Department of Engineering developed a mother robot that is able to design and assemble robot children, and that tests them by measuring their performance in a target task in order to reuse the characteristics of the best candidates to design and build the next generation.

In the whole process, neither computer simulation was used nor human intervention, except the initial input to build the first robot child.

Five experiments were conducted and, in each of them, 10 generations of 10 robots were generated and evaluated. Basis for assessment was the distance covered by the children in a given time.

Dr. Luzius Brodbeck, Dr. Simon Hauser, and Dr. Fumiya Iida collected results of the experiments in the article “Morphological Evolution of Physical Robots through Model-Free Phenotype Development”, published last June on the free-access journal PLOS One under a Creative Commons Attribution license.

The developmental process

In the experiment environment, the mother robot had access to a set of cubic modules to put together in order to build the child.

Developmental process

Developmental process. Source: PLOS ONE. License: © 2015 Brodbeck et al. CC-BY

Each child is characterized by its genome, the set of information (genes) which determines its compositions and configuration. Genes contain three kind of information for the mother:

• the type of module to pick (able to move or not);

• the construction process to follow (translations, rotations, assembly);

• the motor control.

After the construction of each child, the mother puts it in the test area and tests its speed. Once a whole generation has been evaluated, the mother creates the genomes for the next generation by using data collected from the previous performances.

Best children’s genomes stay unaltered so as to preserve their traits and abilities. Remaining genomes are modified. This optimization process is based on an evolutionary algorithm, which improves the construction process of the robot children.

Evolution is thus performed through mutation, where components in one gene are modified or single genes are added or deleted, and crossover, where a new genome is formed by merging genes from two individuals.


Robot Child

One of the mother robot’s child. Photograph: Uniiversity of Cambridge

Results of the experiment were amazing: during the process a large variety of shapes was generated, and the last generation proved to be on average twice as fast as the first one, which demonstrated the capability of the mother robot to identify and perpetuate traits of the best candidates.

Following what happens in nature, artificial evolution is leading to machines that don’t evolve according to a model, and that give rise to unforeseeable possible solutions and combinations.

Lead researcher Dr Fumiya Iida stated: “Natural selection is basically reproduction, assessment, reproduction, assessment and so on. That’s essentially what this robot is doing – we can actually watch the improvement and diversification of the species.