Method for evolving the neural network of a robot. Valid gene sequences are extracted (magnifying lens) from a binary string representing the genome of the robot. Those genes are translated into neurons of different types (colors) according to the genetic specifications, such as sensory, motor, excitatory, or inhibitory neurons. The corresponding neural network is connected to the sensors and motors of the robot and the resulting behavior of the robot is measured according to the fitness function. The genomes of the individuals that had the worst performance are discarded from the population (symbolically thrown into a dustbin) whereas the genomes of the best individuals are paired and crossed over with small random mutations to generate new offspring (the process of selective reproduction is symbolically shown to occur in a mother robot). After several generations of selective reproductions with mutations, robots display better or novel behaviors.

Behaviour evolved in the real world on the Sussex gantry robot in 1994. Controllers (evolved neural networks plus visual sampling morphology) are automatically evaluated on the actual robot. The required behaviour is a shape discrimination task: to move to the triangle, while ignoring the rectangle, under very noisy lighting conditions.

Evolved-walking behaviors on an octopod robot. Multiple gaits and obstacle avoidance can be observed. The behavior was evolved in a minimal simulation by Nick Jakobi at Sussex University and is successfully transferred to the real world as is evident from the video.

Evolved, simulated hexapod walks over rough terrain while homing on a beacon. This behavior was incrementally evolved with the controlling neural-network architecture which was expanding at each stage. Work done at Sussex University by Eric Vaughan.

The video shows stages of evolution of bipedal walking in a simulated, bipedal robot using realistic physics (from the work by Torsten Reil and originating at Sussex University). This was the first example of successfully- evolved bipedal gaits produced in a physics-engine-based simulation. The problem is inherently dynamically unstable, thus making it an interesting challenge.

The video shows a successfully evolved GasNet controlling a simulated robot engaged in a visual-discrimination task under noisy lighting. The GasNet architecture and all node properties are evolved along with the visual sampling morphology (parts of the visual field used as inputs to the GasNet). A minimal simulation is used which allows transfer to the real robot (see Sussex gantry Video 371). A highly minimal controller and visual morphology have evolved. The system is highly robust, coping with very noisy conditions. As can be seen, the GasNet employs multiple oscillator subcircuits - partly to filter out noise. Work by Tom Smith and Phil Husbands.

Identical evolved robots are required to coordinate by coming together and moving off in the same direction. No roles are pre-assigned. The robots must evolve to coordinate such that one robot takes on the role of leader and the others follow. Only minimal sensing is available (proximity IR sensing) and no dedicated communication channels. The robot neural-network controllers are evolved using a minimal simualtion and, as can be seen, these successfully transfer to reality. Work by Matt Quinn, Giles Mayley, Linc Smith and Phil Husbands at Sussex University.

This video demonstrates a robot that is able to change its morphology. It is here shown that this change enables evolution to create useful controllers for this robot faster than a comparable robot that does not undergo morphological change.