Explorations in Real-Time Robotics Using Parallel Processing, Neural Networks and Genetic Algorithms

Real-time robot control has always presented researchers with great difficulties in terms of both the accuracy of the command actions required and also the efficiency by which the commands are obtained. A very important characteristic of the new generation of robotic systems is the presence of intelligent capabilities which is being rapidly supported by fast computing power and adequate sensory equipment. A general form of the robot control loop is shown in Fig(1), where the required job is first divided by the task planner producing a number of consecutive tasks, followed by the motion planer, which gives a time history of positions, velocities and accelerations, sufficient and necessary to realise each task. Once the desired motion elements are available, they are used to produce the commands for the individual joint loops via the control module which may or may not include the dynamic model of the system (model reference adaptive controllers vs. simple PID's). The motion is realised by applying the control commands to the robot system and a feedback module is provides the actual motion elements to cater for any uncertainties and/or changes in the system parameters and/or environment set up. Overall intelligence may be needed at different parts of the control loop, e.g. in connection with the task planner, dynamic model or sensory feedback.