New Zealand’s economy relies on primary production to a great extent, where use of the technological advances can have a significant impact on the productivity. Robotics and automation can play a key role in increasing productivity in primary sector, leading to a boost in national economy. This thesis investigates novel methodologies for design, control, and navigation of a mobile robotic platform, aimed for field service applications, specifically in agricultural environments such as orchards to automate the agricultural tasks.

The design process of this robotic platform as a non-holonomic omnidirectional mobile robot, includes an innovative integrated application of CAD, CAM, CAE, and RP for development and manufacturing of the platform. Robot Operating System (ROS) is employed for the optimum embedded software system design and development to enable control, sensing, and navigation of the platform.

3D modelling and simulation of the robotic system is performed through interfacing ROS and Gazebo simulator, aiming for off-line programming, optimal control system design, and system performance analysis. Gazebo simulator provides 3D simulation of the robotic system, sensors, and control interfaces. It also enables simulation of the world environment, allowing the simulated robot to operate in a modelled environment. The model based controller for kinematic control of the non-holonomic omnidirectional platform is tested and validated through experimental results obtained from the simulated and the physical robot.

The challenges of the kinematic model based controller including the mathematical and kinematic singularities are discussed and the solution to enable an optimal kinematic model based controller is presented. The kinematic singularity associated with the non-holonomic omnidirectional robots is solved using a novel fuzzy logic based approach. The proposed approach is successfully validated and tested through the simulation and experimental results.

Development of a reliable localization system is aimed to enable navigation of the platform in GPS-denied environments such as orchards. For this aim, stereo visual odometry (SVO) is considered as the core of the non-GPS localization system. Challenges of SVO are introduced and the SVO accumulative drift is considered as the main challenge to overcome. SVO drift is identified in form of rotational and translational drift. Sensor fusion is employed to improve the SVO rotational drift through the integration of IMU and SVO.

A novel machine learning approach is proposed to improve the SVO translational drift using Neural-Fuzzy system and RBF neural network. The machine learning system is formulated as a drift estimator for each image frame, then correction is applied at that frame to avoid the accumulation of the drift over time. The experimental results and analyses are presented to validate the effectiveness of the methodology in improving the SVO accuracy.

An enhanced SVO is aimed through combination of sensor fusion and machine learning methods to improve the SVO rotational and translational drifts. Furthermore, to achieve a robust non-GPS localization system for the platform, sensor fusion of the wheel odometry and the enhanced SVO is performed to increase the accuracy of the overall system, as well as the robustness of the non-GPS localization system. The experimental results and analyses are conducted to support the methodology.