An agent must locate and identify elements within its environment in order to manipulate them constructively. A population of physically simulated agents is produced with arbitrary evolved morphologies and neural network controllers without wheels or imposed symmetry. This is done via the use of incremental evolution. It is shown that they can locate and approach a coloured object within their environment using a light detecting sensor with a field of view. An evolved agent is examined and is shown to use a close dynamic coupling between the controller and morphology to produce a stable platform for a visual sensor, allowing it to fixate on its target.
展开▼
