Novel Neuro-dynamic Methods For Solving Robotic Planning And Control Problems
MetadataShow full item record
Robotic systems are growing more complex to cater to the tasks they are required to solve. Irrespective of these tasks the problem of planning and control must be solved for all robotic systems. In this dissertation a general description of that problem is given along with the issues that it faces. This problem concerns planning a trajectory in the operational space of the robot, i.e. the space in which a task is to be executed. Then solving the inverse kinematics problem to compute the joint positions that correspond to that trajectory. And lastly, controlling the robot to track those joint trajectories. One of the issues facing the planning and control problem is singular configurations of a robot; these are joint configurations at which the inverse kinematics problem has multiple solutions. They can also induce large velocities in a robot, and lead to damage in hardware. Another issue to be addressed while solving the planning and control problem is model uncertainties, where the model used to control the robot does not account for all possible parameters that affect its motion. The solution to the planning and control problem presented in this research uses tools from artificial intelligence and tackles these issues to provide a robust framework for planning and control that can be implemented on any robotic chain. By using dynamic neural networks with intelligent feedback, adaptive operational space trajectories are generated. Two methods for dealing with singular configurations are presented, one uses artificial potential functions to avoid those configurations while another uses neuro-dynamic solutions to render the system robust to their effects. To control the robot, artificial neural networks with online learning are used in a neuro-adaptive controller. This controller allows for tracking the desired joint values calculated from the inverse kinematics solution.Numerical simulations are used to verify the capabilities of the developed methods for solving the planning and control problem. These methods also tested in a realistic simulation environment of the Atlas humanoid robotic system. In house software is developed to realize an execution of a walking gait using the neuro-dynamic framework. Results show that the developed methods form a singularity robust framework for solving the planning and control problem for a complex robotic system.