It is desirable to have a robot make a motion in the least time. Current methods result in a compromise between computation time and motion time. For example, classical control methods work in real-time, but result in motions with sub-optimal times. However, optimization routines can produce motions with optimal times, but computation might take minutes. The ideal solution is one which works in real time, and finds an optimal motion time.
Neural Networks are powerful computers based on the structure of the human brain. These may be applied to problems like Optimal Time Motion Planning. The Neural Networks are very fast and fault tolerant. Based on a few examples they are capable of learning to solve an un-modelled problem. A majority of this thesis is involved with trying to produce some examples to train the neural network with.
Three cases were examined for this problem. Two simple cases were examined based on a maximum joint velocity first, and then maximum joint acceleration second. Results from both of these cases proved very successful. The thesis then describes a neural network for maximum joint torque control. This description includes techniques for finding paths based on maximum joint torque limits.
To find Optimum Motions, Bezier Splines are used to model the motion. These are then adjusted by combinations of various Optimization techniques. The final results are good approximations of Optimal Time Motions. These are used to find a number of example motions through the workspace. Sample points are taken from the motions, and these are used to train the neural network.
After training, the Neural Network was tested for various starting and goal points not previously trained. For a given state (position and velocity) and goal input to the Neural Network, the torque output generated a nearly global optimum motion, as compared to the motion generated by the optimization routine.