This thesis describes the development and implementation of mobile robot control techniques that learn to control the execution of a fast and accurate movement whilst being exposed to significant sensory feedback delay. The need for sensing in robotics, such as machine vision, usually incurs significant sampling and processing delays during acquisition (in the order of hundreds of milliseconds), before sensor data can be used for action generation. Typical feedback control techniques can only control a robot using delayed data if the resulting movement is slow enough to allow sensory information to catch up. If a fast movement is required under these conditions, feedback control alone will result in serious instability.
Biological systems such as the human body also suffer from significant sensory delay, yet we are able to perform a staggering array of fast and accurate movements. A part of the brain called the cerebellum is thought to be the mechanism that overcomes this apparent anomaly. The models developed in this thesis are inspired by the biological cerebellum.
The thesis develops models based on a range of theories of cerebellar function. The first class of models work on the theory that the cerebellum is a command generator/modulator, where implicitly predictive commands are learnt for a sequence of delayed state observations. The second type of model explores the idea of the cerebellum acting as a state predictor, where current state predictions are learnt either in the form of a forward model or in response to the presentation of delayed state information. Other techniques which combine both command generator and state predictor theories, such as paired forward/inverse models in a sensory delayed system, were also investigated.
The thesis reviews both biologically and computationally inspired cerebellar literature, and pays particular attention to the performance of various theories under the constraint of sensory delay. It was found that additional mechanisms were required with the various models to accommodate the detrimental effects associated with sensory delay. Rather than being able to function in the presence of sensory delay (to be biologically plausible), the thesis promotes the idea that the cerebellum is used primarily to overcome the effects of sensory delay.
The models developed within, first and foremost, provide a reliable solution to an engineering problem. By using the cerebellum as inspiration, much of what was discovered could be related back to biology and used to provide an example of what would be required in a biologically plausible model. In this way the thesis combines both biology and engineering in an effort to further both domains.