Intelligent production scheduling has been viewed as a problem-solving task that involves the generation of a suitable schedule among different ones using learning techniques. This work presents a unified model for a class of intelligent machines, suitable for flexible manufacturing systems, in which each alternative production plan comprises a fixed number of sequential steps, and the task is to select the optimal plan as more experience is obtained during the operation of the system. This model, based on the theory of hierarchically intelligent control systems developed by Saridis (1995), combines the powerful high-level decision making with advanced mathematical modeling and synthesis techniques of system theory and methods of dealing with imprecise or incomplete environment information.

This paper introduces a design methodology for intelligent controllers, based on a hierarchical linguistic model of command translation by tasks-primitive tasks-primitive actions, and on a two-stage hierarchical learning stochastic automaton that models the translation interfaces of a three-level hierarchical intelligent controller. The methodology relies on the designer's a priori knowledge on how to implement by primitive actions the different primitive tasks which define the intelligent controller. A cost function applicable to any primitive task is introduced and used to learn on-line the optimal choices from the corresponding predesigned sets of primitive actions. The same concept applies to the optimal tasks for each command, whose choice is based on conflict sets of stochastic grammar productions. Optional designs can be compared using this performance measure. A particular design evolves towards the command translation (by tasks-primitive tasks-primitive actions) that minimizes the cost function.

A model for the organization level of a class of intelligent machines suitable for industrial applications is established. This model based on the theory of hierarchically intelligent control systems developed by Saridis (1983) combines the powerful high level decision making with advanced mathematical modeling and synthesis techniques of system theory and methods of dealing with imprecise or incomplete environment information.

A new metric for performance assessment of intelligent machines has been developed. The method fuses concepts from the Theory of Intelligent Machines proposed by Saridis (1988) with traditional reliability analysis in the development of a measure which reflects both the uncertainty inherent in the intelligent machine and the uncertainty allowed by the task description. The metric is entropy based, and is shown to be analogous to a measure of system reliability.

A methodology for performance improvement of intelligent machines based on hierarchical reinforcement learning is introduced. Machine decision making and learning are based on a cost function which includes reliability and a computational cost of algorithms at the three levels of the hierarchy proposed by Saridis. Despite this particular formalization, the methodology intends to be sufficiently general to encompass different types of architectures and applications. Novel contributions of this work include the definition of a cost function combining reliability and complexity, recursively improved through feedback, a hierarchical reinforcement learning and decision making algorithm which uses that cost function, and a methodology supported on information-based complexity for joint measure of algorithm cost and reliability. Results of simulations show the application of the formalism to intelligent robotic systems.>

Intelligent machines are defined to be hierarchically intelligent control systems composed of three levels: the organization level, the coordination level, and the execution level. The authors present an analytical model for the coordination level of intelligent machines, which, together with the established mathematical formulation for the organization level and the well-developed control theory for the execution level, completes the first step toward a mathematical theory for intelligent machines. The framework of the coordination level is a tree structure consisting of a dispatcher and a number of coordinators. A new type of transducer, Petri net transducers (PNTs), is introduced to serve as the basic module in the analytical model. The PNTs can be used to represent the concurrence and conflict among the individual processes within the dispatcher and coordinators. Coordination structures are introduced as a formalism for the specification of integration in the coordination level. The task precedence relationship in the coordination process is presented by the Petri nets derived from the coordination structures. A case study is presented.>

A formulation that makes possible the integration of collision prediction and avoidance stages for mobile robots moving in general terrains containing moving obstacles is presented. A dynamic model of the mobile robot and the dynamic constraints are derived. Collision avoidance is guaranteed if the distance between the robot and a moving obstacle is nonzero. A nominal trajectory is assumed to be known from off-line planning. The main idea is to change the velocity along the nominal trajectory so that collisions are avoided. A feedback control is developed and local asymptotic stability is proved if the velocity of the moving obstacle is bounded. Simulation results verify the value of the proposed strategy.>

The theory of intelligent machines proposes a hierarchical organization for the functions of an autonomous robot based on the principle of increasing precision with decreasing intelligence. An analytic formulation of this theory using information-theoretic measures of uncertainty for each level of the intelligent machine has been developed. The authors present a computer architecture that implements the lower two levels of the intelligent machine. The architecture supports an event-driven programming paradigm that is independent of the underlying computer architecture and operating system. Execution-level controllers for motion and vision systems are briefly addressed, as well as the Petri net transducer software used to implement coordination-level functions. A case study illustrates how this computer architecture integrates real-time and higher-level control of manipulator and vision systems.>

Petri nets are a well established mod- elling technique for analyzing parallel systems. When coupled with an event-driven operatine system, Petri nets can provide an effective means for integrating and controlling the functions of distributed robotic appli- cations. Recent work has shown that Petri net graphs can also serve as remarkably intuitive operator inter- faces. In this paper the advantages for using Petri net as high-level controllers to coordinate robotic functions will be outlined; the considerations for designing Petri net controllers will be discussed; and simple Petri net structures for implementing an interface for operator supervision will be presented. I. INTRODUCTION The Theory of Intelligent Machines, as proposed in (8), describes a hierarchical organization of the functions of an autonomous robot into three levels: an Execution Level containing the hardware and basic control functions, a Coordination Level integrating the capabilities of the in- telligent machine (IM) across hardware systems, and an Organization Level providing higher-level planning and reasoning capabilities. Figure 1 illustrates this hierarchi- cal organization. An analytic formulation of the theory using information-theoretic measures of uncertainty for each level of the intelligent machine has been developed in recent years 9)(10). A long-term research goal of the ration (CIRSSE) is to attempt, guided by these theories, implementations of autonomous robots. In the past year at CIRSSE a computer architecture was developed that implements the lower two levels of the in- telligent machine (4). The architecture supports an event- Center for Intel I igent Robotic Systems for Space Explc-

A control strategy for real-time collision avoidance of a mobile robot in an environment containing moving obstacles is proposed. Objects, including the robot, are modelled as convex polyhedra. Collision avoidance is guaranteed if the minimum distance between the robot and the objects is nonzero. A nominal trajectory is assumed to be known from offline planning. The main idea is to change the velocity along the nominal trajectory so that collisions are avoided. Consistency with the nominal plan is desirable. The process is formulated as an optimization problem and a close to optimal solution is obtained.>

Given an explicit task to be executed, an intelligent machine must be able to select a plan containing control and sensing strategies such that a desired set of specifications are reliably satisfied. This issue is addressed through a technique termed reliable control and sensor fusion. An approach to reliability analysis is proposed based on the use of entropy which is invariant with respect to homogeneous coordinate frame transformations. This yields an easily calculated set of entropy constraints that must be met to ensure reliable operation. An example problem illustrates the technique and presents general results for visual positioning.>

An analytical model for the coordination level (CL) of intelligent machines is presented. The framework of the CL is a tree structure consisting of a dispatcher and a number of coordinators. Petri net transducers have been used as the basic module in the theory to describe individual task translating processes of the dispatcher and coordinators. Coordination structures are introduced as a formalism for the specification of integration in the CL. The task precedence relationship in the coordination process is presented by the Petri nets derived from the coordination structures. These Petri nets also provide a formal approach for using the concepts and analysis methods in Petri net theory to investigate the properties of the coordination structures.>

The dynamic equations of motion are presented for two or more cooperating manipulators on a freely moving mobile platform. The system of cooperating robot manipulators forms a closed kinematic chain where the force of interaction must be included in the formulation of robot and platform dynamics. The formulation includes the full dynamic interactions from arms to platform and arm tip to arm tip, and the possible translation and rotation of the platform. The equations of motion are shown to be identical in structure to the fixed-platform cooperative manipulator dynamics. The structure of the closed-chain dynamics, allows the use of any solution for the open topological tree of base and manipulator links. The number of degrees of freedom (DOF) of the system is sufficiently large to make recursive dynamic calculation methods potentially more efficient than closed-form solutions. A complete simulation with two 6-DOF manipulators of a free-floating platform is presented along a with a multiple-arm controller to position the common load.>

A probabilistic approach is followed for the modeling of intelligent robotic systems composed of three interactive levels of organization, coordination, and execution of tasks. Probability and entropy functions are used as analytic measures to interpret mathematically the system functions. The modeling constraint obeys the mathematically proven principle of increasing precision with decreasing intelligence. This probabilistic approach has been tested by simulation in connection with overcoming emergency situations in a pressurized water reactor nuclear power plant.>

A Petri net model of the coordination level of an intelligent mobile robot system (IMRS) is presented. The purpose of this model is to specify the integration of the individual efforts on path planning, supervisory motion control, and vision system that are necessary for the autonomous operation of a mobile robot in a structured dynamic environment. This is achieved by analytically modeling the various units of the system as Petri net transducers and explicitly representing the task precedence and information dependence among them. The model can be used to simulate the task processing and evaluate the efficiency of operations and the responsibility of decisions in the coordination level of the intelligent mobile robot system. Some simulation results of the task processing and learning are presented.>

By using concepts from information theory and reliability theory, new techniques are proposed for finding the reliability corresponding to alternative subsets of control and sensing strategies such that a desired set of specifications can be satisfied. The analysis is straightforward provided that a set of Gaussian random state variables is available. An example problem illustrates the technique and provides general reliability results for visual serving with a computed torque control algorithm.>

Robotic tasks in space require manipulation of massive objects capable of attaining large momentum. The momentum can pose hazardous conditions and introduce destabilizing effects on a space platform. Consequently, a technique for limiting the momentum applied to objects under manipulation subject to arbitrary velocity input commands is proposed. The algorithm does not require mass position or inertia information about the object, and it takes actuator limitations into account in forming the momentum limits. To evaluate the probability that a velocity trajectory will fall within the momentum bounds, reliability theory is employed. This enables autonomously generated trajectories to be validated for compliance with momentum limits.>

A recursive formulation is presented for the calculation of the inverse and forward dynamics of rigid robot manipulators with gear systems on each joint. The complete effects of the gear ratios and the gyroscopic effects of the spinning motor/gear are included in the recursive formulation. The forward dynamics solution recursively calculates the joint accelerations when given motor torques, and the number of computations grows linearly with the number of links.>

The dynamic equations of motion for two manipulators holding a common object on a freely moving mobile platform are developed. The full dynamic interactions from arms to platform and arm-tip to arm-tip are included in the formulation. The development of the closed chain dynamics allows for the use of any solution for the open topological tree of base and manipulator links. In particular, because the system has 18 degrees of freedom, recursive solutions for the dynamic simulation become more promising for efficient calculations of the motion. Simulation of the system is accomplished through a MATLAB program, and the response is visualized graphically using the SILMA Cimstation.>

Given an explicit task to be executed, an intelligent machine must be able to find the probability of success, or reliability, of alternative control and sensing strategies. By using concepts for information theory and reliability theory, new techniques for finding the reliability corresponding to alternative subsets of control and sensing strategies are proposed such that a desired set of specifications can be satisfied. The analysis is straightforward, provided that a set of Gaussian random state variables is available. An example problem illustrates the technique, and general reliability results are presented for visual servoing with a computed torque-control algorithm. Moreover, the example illustrates the principle of increasing precision with decreasing intelligence at the execution level of an intelligent machine.>

A three-tier structure consisting of organization, coordination, and execution levels forms the architecture of an intelligent machine using the principle of increasing precision with decreasing intelligence from a hierarchically intelligent control. This system has been formulated as a probabilistic model, where uncertainty and imprecision can be expressed in terms of entropies. The optimal strategy for decision planning and task execution by the intelligent machine can be found by minimizing the total entropy in the system. The focus is on the design of the organization level of the intelligent machine as a Boltzmann machine, as described in current neural network literature. Since this level is responsible for planning the actions of the machine, the problem at this tier is formulated as the construction of the right sequence of tasks or events that minimizes the entropy for the desired action. The Boltzmann machine is reformulated to use entropy as the cost function to be minimized. Simulated annealing, expanding subinterval random search, and the genetic algorithm are presented as search techniques to efficiently find the desired action sequence. The genetic algorithm is modified to ensure convergence in probability to the minimum entropy of the system. Simulations of these algorithms are shown and performance is analyzed.>

The first step in the creation of a robotic manipulator controller performance database has been completed. Experimental evaluation of the trajectory-tracking performance of a PUM-600 under computed-torque control has provided valuable engineering insight into the role of dynamics compensation for manipulators with high torque amplification drive systems. Experimental results validate the assumptions that Coriolis and centrifugal forces are negligible and actuator dynamics play a central role in manipulator dynamics. Without complete modeling of the drive-system dynamics, computed-torque trajectory-tracking accuracy was inadequate for gross motion control. It is concluded that more accurate representation of the real robot system dynamics is required for realistic simulation of modern control algorithms and improved real-time performance.>

A simple method of trajectory generation of robot manipulators is presented. It is based on an optimal control problem formulation. The jerk, the third derivative of position, of the desired trajectory, adversely affect the efficiency of the control algorithms and therefore should be minimized. assuming joint position, velocity and acceleration to be constrained, a cost criterion containing jerk is considered. Initially, the simple environment without obstacles and constrained by the physical limitations of the joint angles only is examined. For practical reasons, the free execution time has been used to handle the velocity and acceleration constraints instead of the complete bounded state variable formulation. The problem of minimizing the jerk along an arbitrary Cartesian trajectory is formulated and given analytical solution, making this method useful for real-world environments containing obstacles.>

A hybrid position/force control algorithm was implemented on a PUMA 600 series robot. Multiple processors were used to increase the sampling frequency of the control algorithm. The force control algorithm used the hand Jacobian and hand frame force measurement fingers mounted in a gripper to control the forces and moments applied by the end effector. The position control algorithm used the Lagrange-Euler formulation of the dynamics of the arm to control the joint space trajectory of the arm. The algorithms' performance was evaluated using a proportional controller for the force control half and a proportional-derivative controller for the position control half. Several programs were written to show various force and position tracking abilities.>

Generalized N-dimensional information theory is utilized to describe quantitatively the flow of knowledge (information) within intelligent robotic systems modeled within the constraints imposed by the theory of intelligent controls. A generalized partition law of information rates is derived to describe analytically the laws of information which govern the performance of the three interactive levels of such systems as well as within each of the higher two levels, namely the organization and coordination levels. The law considers the possible trade-offs between the noise, the internal control procedures, the feedback mechanisms, possible coordination, and memory exchange procedures generated during the operation of intelligent robotics systems.>

Forces unmodeled by Lagrange-Euler dynamics restrict the trajectory tracking accuracy of dynamics based PUMA manipulator control laws. The ability of unmodeled force compensation techniques to eliminate those restrictions, enabling dynamics based PUMA controllers to accurately track a high speed trajectory, has been identified. Implementation of nonlinear velocity dependent friction compensation in the feedforward loop significantly improves controller efficacy. A combination of feedforward and feedback compensation techniques permits tracking accuracy sufficient for gross motion control of a PUMA manipulator without additional instrumentation. The effectiveness of proposed modern control theories can now be compared to an experimentally determined manipulator control performance baseline.

Evaluation of the effects of dynamics on real-time closed loop robotic manipulator joint control requires accurate, computationally efficient dynamics algorithms. Through a study of open loop joint torques the dominant and negligible torque components for each link of a PUMA-600 have been identified so that prudent model reduction could be accomplished. An acceleration related matrix model which requires only 94 multiplications and 59 additions while achieving accuracy to within ten percent of the complete model has been developed and justified. Reduction of the Newton-Euler PUMA- 600 dynamics to 304 multiplications and 212 additions has also been achieved without any loss of accuracy. The developed algorithms support our real-time study of the effects of dynamics on robot control.

A Hierarchical Control System with organizer, coordinator and hardware levels has been developed. A VAX-11/750 serves as the organizer. An LSI-11/23 equipped with FPF11 floating point processor is the basis of the coordinator. The original controller for a PUMA-600 robotic manipulator has been modified to serve as the hardware level. Organizer level software has been created for communication and control of the coordinator. Coordinator command and control software development has duplicated and expanded upon the functions of VAL. Real-time evaluation of modern robotic manipulator control methods on a PUMA-6O0 is now possible.

The real-time and simulation performance of the computed-torque joint control technique employing four configurations of dynamic models in controlling a six degree of freedom robotic manipulator has been thoroughly evaluated over a set of twelve different operational situations. Simulation comparisons have been conducted to predict the repercussions of neglected dynamics on computed-torque performance. Real-time tests have been conducted using a PUMA-600 and modified controller. Real-time results do not adhere to the simulation predictions. The algorithm with the simplest dynamics performs best in real-time.

A mathematical formulation of the organization level of an intelligent machine, modeled as a Knowledge Based (Expert) System is derived and analyzed. Characteristic functions of the organization level of such a machine are: machine reasoning, machine planning, decision making, feedback (learning) and memory exchange. Probabilistic models are used to express analytically their individual functions. Information Theory measures and concepts are utilized to describe quantitatively the flow of knowledge in that level as the machine operates. An example illustrates this approach.

A ranging method is presented that is being developed at the Robotics and Automation Laboratory at R.P.I. The method can be used to locate objects of different shape in 3-D space. A camera at a fixed position is observing a scene and a light source attached to a robot arm sending a beam of light to the scene represent the hardware used. The position of the camera and the light source along with the picture of the scene provide enough information to compute the reflection point of the light beam in the scene. An error analysis of the process, the robot arm and the camera system are presented and the results of the implementation discussed.

An optimal control formulation for robotic manipulators is proposed in this paper. The Computed Torque Control, Resolved Acceleration Control, and a PID End Point Control are derived as special cases of this formulation. Torque and acceleration feedback are adopted to eliminate the on-line calculations of feed forward components in the above control algorithms, and to increase robustness of the control system. This controller can be implemented approximately without torque sensors and accelerometers. The resulting control system is proved to be bounded input bounded state stable. Simulation results on both the Scheinman MIT and PUMA 600 arms are presented.

Design modifications on the controller of the PUMA- 600 arm are required to apply modern control techniques in real-time. These modifications are performed on the computer-level and the joint control level of the PUMA controller. Real-time closed loop control of the arm is thus made possible. The paper compares the performance of various control algorithms in real-time and by simulation.

An explicit solution for the robot arm controller, optimal in the linear quadratic sense, is obtained. Based on this formulation, a PID controller with acceleration feedback is proposed which is computationally efficient and robust relative to variations in the dynamic model and external disturbances.

Using the equivalence of the average Lagrangian to an Entropy established in Theoretical Thermodynamics, this paper expresses the average value of the performance criterion of a feedback control problem as an entropy. Thus, the optimal control problem may be recast as an information theoretic, one which minimizes the entropy of selecting the feedback controls. This unifies the treatment of all the levels of a Hierarchically Intelligent Control System by minimizing the sum of their entropies. Such a system is composed of three levels hierarchically ordered in decreasing intelligence with increasing precision: the organization level, performing information processing tasks like planning, decision making, learning and storage and retrieval of information from a long-term memory; the coordination level, dealing again with information processing tasks like learning, lower level decision making and dealing with short term memory only and the control level, which performs the execution of various tasks through hardware using feedback control methods. A mathematical programing algorithm which utilizes one step ahead entropies may be used to approximate the optimal solution of the intelligent machine problem.

The development of a new hierarchical linguistic-based learning control structure for complex systems is described. The complete system will be able to interact with a human operator in a high-level language at the highest level of the hierarchy and will be able to control detailed motions of a complex physical system at the lowest level of the hierarchy. Each level of the hierarchy is defined by a formal grammar which can generate exactly the class of admissible control actions at that level. A new linguistic structure, the linguistic decision schema, is proposed to specify the mapping between linguistic elements in adjacent levels. In the most general form, the decision schema incorporates a learning algorithm to obtain asymptotically optimal mappings for control under stochastic environments.

This work describes a system developed on a PDP 11/45 computer which recognizes spoken commands and controls a Scheinmann MIT mechanical arm to serve the needs of an immobilized patient. The speech recognition is discrete word with template matching, using as features zero crossing rate, average absolute magnitude, and frequency and normalized error as derived from the two pole linear predictive analysis. Commands are made up of words from a 16-word vocabulary obeying a specified, syntax. Upon recognition of a command the arm performs the corresponding pre-programmed task. A recognition time of two seconds is attained (including the one second sampling interval), plus an additional one second to verify the word. With the command syntax a recognition rate of 88% is attained, as compared with only 56% when each word is matched against every word in the vocabulary.

A theoretical procedure is developed for comparing the performance of arbitrarily selected admissible controls among themselves and with the optimal solution of a nonlinear optimal control problem. A recursive algorithm is proposed for sequential improvement of the control law which converges to the optimal. It is based on the monotonicity between the changes of the Hamiltonian and the value functions proposed by Rekasius, and may provide a procedure for selecting effective controls for nonlinear systems. The approach has been applied to the approximately optimal control of a trainable manipulator with seven degrees of freedom, where the controller is used for motion coordination and optimal execution of object-handling tasks.

A hierarchical method combining analytical techniques from control theory and heuristic techniques from artificial intelligence is presented and applied to the decentralized control of a prosthetic arm. The dynamic model of the arm is derived, and two complementary performance criteria are suggested for the kinematic and the dynamic evaluation for the system response. The "principle of minimum interaction" is used to decompose the prosthetic system into seven subsystems, one per mechanical degree of freedom. A "suboptimal" control structure for nonlinear systems is proposed in conjunction with a performance adaptive self-organizing control algorithm. Syntactic pattern classification is used for the dynamic coordination of the subsystems. The syntax of the man-machine commands is also examined as part of the function of highest level of the hierarchy, the organizer.

A fundamental problem in system modeling and theory is the characterization of the structure of an unknown nonlinear stochastic system when only input-output measurements are available. A method of classifying nonlinear stochastic systems, using pattern recognition and a pattern vector constructed from the input-output data, is proposed for ten stated classes of low-order nonlinear systems. The method is capable of extension to additional classes of nonlinear systems. Extensive experimental results are given to show that classification of an unknown nonlinear system, with respect to basic structural properties, can be and accomplished with a very high probability of correct classification. Various applications of the classification procedure are given, particularly in the areas of systems modeling, self-organizing control systems, and learning control systems.

The multidimensional Kiefer-Wolfowitz stochastic approximation algorithm is proposed for the performance-adaptive self-organizing control of a class of distributed systems. The class of systems considered is that which can be modeled mathematically by the general linear second-order elliptic equation with unknown coefficients and are defined on a fixed and bounded domain Ω. Noisy measurements of the response are assumed available at a finite number of fixed points in Ω. The minimum number of measurement points necessary for recovery of the significant part of the response is determined by application of an n-dimensional extension of the sampling theorem.