Learning visuo-motor behaviours for robot locomotion over difficult terrain

This thesis investigated efficient methods for learning locomotion behaviours for robots, and the challenges of combining several complex controllers. Experiments were performed with a dynamic biped in simulation required to walk across gaps, over steps and stairs, and jump over hurdles and blocks, and also in a real-world scenario with a large tracked platform negotiating small doorways. The developed solutions utilised perception to perform complex maneuvers while minimising retraining for new behaviours. Ideas from this thesis lead toward scalable behaviour libraries to enable robots to make their way into an increasing number of roles in our society