This thesis describes the development and implementation of a visually guided robot that learns to perform a task best described as "hunt and gather". The hunt and gather operation involves the negotiation of an environment in search of particular objects (hunting), followed by a retrieval of those objects (gathering). This behaviour forms the basis of a wide variety of dirty, dull and dangerous jobs that would be ideal to assign to robots; tasks such as fruit picking, weeding, domestic or industrial cleanup, or planetary exploration.
CORGI, the robot described in this thesis, learns how to perform a simple hunt and gather operation: finding tennis balls in my office. CORGI has only one sensor, a CCD camera that provides visual information. Using neural network techniques, both perception and control systems are developed, trained and implemented on the robot platform. The resultant system is capable of performing the hunt and gather task in real time: using only a single $20 microprocessor.
The thesis reviews recent literature in robotics and neural networks in the context of Braitenberg's book Vehicles: An Experiment in Synthetic Psychology. The literature review shows that many of Braitenberg's gestalt experiments are practicable to investigate with recent developments in robotic and neural network technology. A simulation, SimCORGI, is used to show that the simplest of networks and training algorithms used in neural network research can be used to train control systems for Braitenberg vehicles. Furthermore, a perception system that is trained to detect drink cans is demonstrated in a pilot study to be suitable as a sensor system for a Braitenberg vehicle.
CORGI was developed based on the findings from these pilot studies. The robot has a single Motorola MC68HC 1 6 microcontroller to perform all of the control, vision processing, communication and peripheral tasks required of the project. The robot's perception system was developed by transmitting images from the robot to another computer, where the images were tagged for supervised training based on the back propagation algorithm. Investigations related to network architecture, and data representation and presentation led to the development of perception systems suitable for the hunt and gather operation.
The trained perception networks were then loaded back to robot for real time non-learning perception. A behaviour based control system was developed based on the principles investigated in simulation and run in real time on board the robot. The resultant system is robust to changes in the environment, and readily generalises its behaviour to novel situations and unseen environments. The robot is also used in a real time behaviour training experiment, in which the robot learns the hunt and gather operation in four minutes of on-board training.
Finally, the thesis reviews the approach to the problem and various aspects of the implementation. Further improvements and extensions to many aspects of the project are suggested.