Karl Mathia

Adaptive semi-autonomous robotic neurocontroller

We have designed a neural network semiautonomous robotic arm controller. This controller performs end-effector path planning, inverse kinematics, and joint control to move the end- effector to a commanded position. We have tested the adaptive neural joint controller and inverse kinematics in simulation. The joint controller has been tested on two real arms. These real arms are the Extendable Stiff Arm Manipulator (ESAM) and the Proto-Flight Manipulator Arm (PFMA). Both of these arms are very different, yet the same unmodified joint controller software can control them both. The controller has also shown tremendous adaptability to large payload variations. It has been shown to adapt to a 35 pound end-effector payload on the ESAM from a zeroed initial state. This ability to handle different arms and payloads is due to the fact that the controller makes no assumptions as to the arms dynamics or payload. The same tests performed on a decentralized PD controller showed that the neural network controller is superior.

Real-time geometrical approximation of flexible structures using neural networks

This study demonstrates the potential of artificial neural networks for the geometrical approximation of flexible structures. Online modeling of the deformation and dynamics of flexible structures can improve the control and performance of systems such as airplane wings, rotor blades of helicopters, large articulated space structures, and robots with flexible links or joints. Here a neural model that approximates the deflection of such structures is developed. Real-time modeling is provided by a specialized neural network processor. We demonstrate this concept using the model for the nonlinear deflection of an viscoelastic airplane wing.

Physics Based Simulation of Light Sources

Light sources inherent in urban populations, battlefield effects, and vehicles are the dominant features in a simulated night scene. The effectiveness of optical sensors in these environments is positively or negatively affected depending upon the quantity, type, power, placement, housing, and orientation of these sources. Yet, traditional simulation techniques have focused little on the accurate representation of light sources and their interaction with other scene elements. Traditional lightpoint rendering techniques, while addressing modest out-the-window training requirements, fall well short of recreating the challenges associated with the employment of image intensifying sensors for situational awareness in today’s complex mission environments. The elevated priority of low-altitude, urban missions such as close air support (CAS) and combat search and rescue (CSAR) will require richer, more accurate, and more dense representations of light sources if the sensor-related elements of these missions are to be trained in simulators.

Control of formations under persistent disturbances

We study the distributed control of autonomous second order agents under persistent disturbances. We show that the usual averaging rule for convergence to formation is only able to reject constant disturbances that are identical for each agent. We also prove that using a distributed dynamic compensation law the system can be made to converge to formation under constant perturbations of the control input even when the perturbations are different for each agent. We illustrate the results with numerical simulations.

Application of highly parallel computing hardware to pattern recognition problems

Neural networks are well known for their ability to perform pattern recognition tasks. This paper discusses the use of parallel neural network hardware for performing pattern recognition tasks. We address the need for neural network hardware and how it can dramatically improve system performance both in training and in actual applications. The use of specialized parallel processing hardware is discussed as well as alternative hardware and software approaches. Finally we give some comparisons between multi-processor computer architecture, Pentium class microcomputers and custom hardware.

A digital processor for neural networks

A special purpose digital processor which is optimized for neural network applications is described. The design is built around a small number of native instructions, each of which correspond to a high level neural network construct, thereby facilitating a highly pipelined processor implementation while simultaneously minimizing the number of instructions required to realize a given network. The neural network processor (NNP(R)) has been implemented for use in both PC/ISA and VME systems and is supported by a full suite of development software.