Neuroevolution of Feedback Control for Object Manipulation by 3D Agents

It has been shown that manipulation of objects by 3D virtual creatures can play an important role in the evolution of complex, embodied sensorimotor behaviours. In this work we examine the capacity of virtual creatures that use evolutionary and control architectures already shown to be capable of sensor-differential gradient-following locomotion (tropotaxis) to adapt to solve a physical problem involving the manipulation of 3D objects in their environments. Specifically, the creatures task is to guide a physically-modelled cube through their environments in order to achieve maximum covered distance of the object. Agents were evolved in the manipulation environment from random initial genotypes and from genotypes previously optimised for performance in a different task. Performance was evaluated both before and after evolutionary adaptation. We show that the architecture achieves embodied feedback control in the block movement task. We observed some overlap between the earlier and later environments but also that success in the first environment does not preclude or entail success in the second. We found that species evolving from scratch do no better or worse than those optimised for a different environment, and that sensory feedback is necessary for correct approach and control behaviours in agents, although close control is less dependent on sensory input than distance approach.