Navigation : EXPO21XX > AUTOMATION 21XX > H05: Universities and Research in Robotics > University of New Hampshire

Company Profile

The research emphasis of the Robotics Laboratory in the Department of Electrical and Computer Engineering is the application of fast associative memories and other neural network learning techniques (such as CMAC neural networks) to problems in control, pattern recognition, and signal processing. The basic concept is to design hardware/software systems which improve their own performance through practice. Details of specific research can be found in published papers and graduate student theses. The Robotics Laboratory currently maintains six experimental settings for research in learning control. The first includes a General Electric P-5 five axis articulated industrial robotic arm which has been the basis of much of our real time experimentation. This arm has been used both for studies of learning high speed dynamics and of learning low speed hand-eye coordination (using video feedback). The second experimental preparation includes two Scorbot-ER V table top robotic manipulators for use in experiments involving path planning, multiarm cooperation and workspace obstacle avoidance. The experiment includes a true binocular vision system which can be positioned and oriented actively using a third table top robotic arm with six degrees of freedom. The third major experimental preparation includes a ten degree of freedom biped walking structure, with force sensing feet and a two-axis accelerometer for sensing balance. The fourth experimental preparation includes a twenty degree of freedom quadruped walking structure, also with force sensing feet and a two-axis accelerometer for sensing balance. The fifth experiment involves a wheeled mobile robot with an array of ultrasonic range finders for studies of adaptive navigation and map building. Finally, the six experiment involves using neural network learning in the myoelectric control path of a Liberty Technology Boston Elbow. Computing in the laboratory is performed primarily using several 80486, Pentium and Pentium-Pro (P6) based engineering workstations, two massively parallel SIMD processors, INMOS transputer based multi-processing systems, and special purpose neural network hardware (developed at UNH). These systems support real time control experiments, simulation studies, general purpose graphics and document preparation. The laboratory also maintains equipment and tools for electronic hardware development and testing (oscilloscopes, signal generators, power supplies, etc.). Vibration Control by Gordon Kraft Nearly everyone has experienced the annoyance of a long drive with an unbalanced tire, or the whir of a noisy hard drive, or seen the blur in a picture taken from a camera that moved as the shutter closed. If you saw the movie \"Hunt for Red October\", you know how important submarine underwater vibrations are to the Navy. The Hubbell telescope cannot function if the supporting platform in space is moving. Factory workers are less efficient if they feel machinery vibrations for long periods of time. All of these are examples of unwanted vibrations. Control of these unwanted vibrations is a very important problem. Most vibration control systems are passive. The rubber in your car engine mounts or in air conditioning ducts are examples. These are called vibration absorbers and by far most vibration reduction systems use these passive elements. In some systems it\'s important to reduce the vibrations beyond the capability of the passive systems. In these cases, an active feedback control loop is required. Simply stated, the vibration measurements taken from various types of acceleration sensors are processed and then applied to dynamic actuators to apply forces that oppose the vibrations. These systems are usually subject to \"ad hoc\" adjustments to tune the feedback controller for the particular application. Once the system is tuned the performance of the system remains fixed. That is, it never gets any better as time goes on. In 1997 we received an NSF grant for $365,000.00 to apply a neural network called CMAC to the area of vibration control. The NSF grant is from the Knowledge Modeling and Computational Intelligence Division of NSF (director is Dr. Paul Werbos). The UNH Robotics Lab version of CMAC has been very successful at other types of control systems such as robotics and signal processing applications. It has advantages over other neural networks such as reduced memory requirements, faster training times and faster real-time control cycles. The key point is that, with CMAC in the control loop, the control system performance will continue to improve with time. As the network accumulates more experience about the system, it is able to continuously improve the control signal to reduce the vibrations more effectively. The network is capable of working with linear or non-linear systems and adjusts itself to changing parameters in the system.