Keele Research Repository

Bipedal walking is a difficult behaviour to encode into an evolutionary neural network, particularly in three-dimensional environments. Agents must be constantly maintaining balance alongside their primary objectives. Here we re-implement a simple evolutionary bipedal system, achieving high fitness and stepping gaits in 3D without the preliminary 2D bootstrapping process required by the original work. This high-performing system, with its deliberately simple neurocontroller, provides an excellent foundation for the community to use for the evolution or learning of more complex behaviours in bipeds. We also investigate the effects of modified morphology with the system, significantly improving agent fitness by evolving networks alongside morphologies resembling a baby albatross. The agents with albatross morphologies travel up to three times further than default agents. We then test incrementally evolving agent morphology via the simultaneous evolution of a separate morphological genotype. We initialised this genotype either alongside a high-performing controller or from a completely random point in both fitness landscapes. Agents evolved from this random initialisation travel up to four times further than default agents. One randomly initialised incremental morphology also achieves gaits with significantly higher upper body and swing knee controller input weights than the default.