NASA is studying a Mars Rover Sample Return (MRSR) mission to perform in-situ analysis, collection and return to Earth of Martian surface samples. The value of science return from this mission is critically dependent upon the ability of a robotic roving vehicle to negotiate the diverse geology of this planet without incurring accidental damage or vehicle entrapment. Legged locomotion offers the considerable advantages of stability, low power, and traversability over extremely rugged terrain, and legged vehicle design concepts are currently being developed under the MRSR1 and Pathfinder2 projects. Semi-autonomous operation of a walking planetary rover entails several unique technical challenges, and places a premium on the system architecture needed to coordinate and control the vehicle actions. A design framework for such a system is provided in the following paragraphs, and is intended for missions where a high degree of autonomy is dictated. It provides a logical computing architecture for rover mobility and local navigation subsystem design by defining a set of functional modules and interfaces to facilitate software and hardware specification.
A flexible robotic assembly cell will have to deal with error detection, diagnosis and recovery. Much of the literature in this area to date has concentrated on devising specific algorithms for detection and recovery, and has largely ignored the central issues of (1) determining what needs to be represented in order to deal with the problem of detecting and recovering from errors and (2) building an appropriate general representation. We begin our work by defining the components of the error detection and recovery problem and by defining what kinds of errors can reasonably be detected. We then construct a generally applicable representation for error detection and recovery information that supports these definitions. The representation, which is based on our previous work on task grammars16, is similar to Donald's3 notation of an EDR plan. We use RS, a process-based model for representing task plans11, as the underlying 'vocabulary' to describe our representation concisely. We evaluate our representation by applying it to the following problems: 1. Distinguishing between planned decision-making and error detection so that the two are not intermixed arbitrarily in a plan. 2. Developing a hierarchical structuring to allow localized error detection and recovery. 3. Gracefully handling errors that cannot be diagnosed. 4. Performing error monitoring on the error monitoring system. Although some formal material is included here, the main emphasis in this paper will be on the representational advantages of our approach.
Controlling robots in dynamic environments poses several unique problems not found in other environments and disciplines. In this paper, we explore some of these problems and detail some of the computational characteristics of processes that are needed to control a robot in such an environment. Having defined these computational characteristics, we then discuss and analyze the system support that would be required in order to execute these processes. Finally we discuss the results obtained using a simulated environment.
Due to sensor errors, uncertainty, incomplete knowledge, and a dynamic world, robot plans will not always be executed exactly as planned. This paper describes an implemented robot planning system that enhances the traditional sense-think-act cycle in ways that allow the robot system monitor its behavior and react in emergencies in real-time. A proposal on how robot systems can completely break away from the traditional three-step cycle is also made.
The models of statistical mechanics provide an alternative to the methods of classical mechanics more traditionally used in robotics. They have a potential to: improve analysis of object collisions; handle kinematic and dynamic contact interactions within the same framework; and reduce the need for perfect deterministic world model information. The statistical mechanics models characterize the state of the system as a probability density function (p.d.f.) whose time evolution is governed by a partial differential equation subject to boundary and initial conditions. The boundary conditions when rigid objects collide reflect the conservation of momentum. The models are being developed to embedd in remote semi-autonomous systems with a need to reason and interact with a multiobject environment.
This paper reviews one method of global kinematic analysis, based on a manifold mapping reformulation of manipulator kinematics, which is suitable for both non-redundant and redundant serial chain manipulators. Two applications of this approach are considered: the relationship between changing pose and singularities for non-redundant manipulators and the homotopy class of redundant manipulator self-motions. The first relationship has important implications for non-redundant manipulator regrasping operations. Variations in redundant manipulator self-motion homotopy can have an important effect on motion and redundancy resolution planning, since they can lead to algorithmic singularities and unexpected loss of manipulator capabilities.
If robots must function in unstructured environments, they must also possess the ability to acquire information and construct appropriate models of the unknown environment. This paper addresses the automatic generation of kinematic models of unknown objects with moveable parts in the environment. If the relative motion between moving parts must be observed and characterized, vision alone cannot suffice. An approach in which manipulation is used with vision for sensing is better suited to the task of determining kinematic properties. In this paper, algorithms for constructing models of unknown mechanical assemblies and characterizing the relative motion are developed. Results of a simulation are described to demonstrate the role of manipulation in such an endeavor.
Current models of robot perception tend to isolate low-level visual processing from behavior and include unrealistic assumptions concerning the modularity of low-level processing. In this paper, I suggest a new way of viewing low-level vision computations based on an extension of Ullman's visual-routines idea. The earliest visual computations are measurements of intrinsic properties of the scene. This is followed by the computation of perceptual organizations. Beyond these two generic types of representations, most visual representations are task-specific and their computations task-dependent. The paper discusses various types of visual measurements that may be useful for a robot-vision system.
A frequent problem in the use of potential functions for robot path planning is that local minima often occur. These local minima may be eliminated by judicious selection of potential functions for goals and obstacles. Specifically, harmonic functions may be used without introducing such minima. While there are analytic, easily superposed solutions for impenetrable point obstacles, this is not the case for impenetrable obstacles with finite, nonzero extent (e.g., walls). Instead, numerical methods that are well suited to massively parallel computation can be used.
A bottleneck in building a knowledge base signal interpretation system is combining the appropriate problem-solving knowledge from the expert with physical observations the signal carry. The information and knowledge processing techniques available in the fields of pattern recognition, signal processing and heuristics can be used to automatically measure the parameters from the physical observations. In this paper, such methods are used to develop an intelligent signal interpretation system. This combined approach will not only automate and accelerate the knowledge acquisition and organization process, but will also formalize and structure the decision making process. The system is designed to interpret and classify signals emitted from a material source. The system consists of four basic components, namely; Fact Gathering, Knowledge Base, Knowledge Formalization, and Inference Engine. The fact gathering subsystem, 1) collects the transduced signals from materials and extracts a large feature set from them, and 2) collects the a priori real-world knowledge and the expert knowledge about the source material and testing conditions. The facts, a priori real-world knowledge, and the pattern measurements (features) are organized into a knowledge base. The next subsystem formalizes the knowledge into a tree structure using a class-association concept. The last subsystem is the Inference Engine which primarily classifies the signals using composite knowledge incorporated in the system. This paper presents the design of the proposed system and shows successful identification of unknown signals from several material defect sources.
What humans actually observe and how they comprehend this information is complex due to Gestalt processes and interaction of context in predicting the course of thinking and enforcing one idea while repressing another. How we extract the knowledge from the scene, what we get from the scene indeed and what we bring from our mechanisms of perception are areas separated by a thin, ill-defined line. The purpose of this paper is to present a system for Representing Knowledge and Recognizing and Interpreting Attention Trailed Entities dubbed as REKRIATE. It will be used as a tool for discovering the underlying principles involved in knowledge representation required for conceptual learning. REKRIATE has some inherited knowledge and is given a vocabulary which is used to form rules for identification of the object. It has various modalities of sensing and has the ability to measure the distance between the objects in the image as well as the similarity between different images of presumably the same object. All sensations received from matrix of different sensors put into an adequate form. The methodology proposed is applicable to not only the pictorial or visual world representation, but to any sensing modality. It is based upon the two premises: a) inseparability of all domains of the world representation including linguistic, as well as those formed by various sensor modalities. and b) representativity of the object at several levels of resolution simultaneously.
This paper describes a system of guidance for an intelligent mobile autonomous system (autonomous robot) based upon an algorithm of "pilot decision making" which incorporates different strategies of operation. Depending on the set of circumstances including the level of "informedness", initial data, concrete environment, and so on, the "personality" of PILOT is being selected between two alternatives: 1) a diligent strategist which tends to explore all available trajectories off-line and be prepared to follow one of them precisely, and 2) a hasty decision maker inclined to make a choice of solution in a rather reckless manner base upon short term alternatives not regarding long term consequences. Simulation shows that these two personalities support each other in a beneficial way.
A new control methodology is presented to effectively operate a robotic system with redundant degrees of freedom. The utilized Decomposed Optimization Technique (DOT) is part of the AISP (An Intelligent Spatial Planner) development. DOT considers the robotic system as several connected subsystems with locally distributed intelligence. Each subsystem has certain degrees of freedom to pursue local optimum state. The resulting parallel distributed processing architecture presents a flexible structure to accommodate sophisticated manipulators with higher level of difficulty. The employed robot dynamics model for each subsystem is generically simple such that the corresponded read-time control scheme can incorporate self-correction mechanism in parameter identification.
Co-ordination of multiple manipulators requires cooperation at several levels in the control hierarchy. A distributed processing environment with no hardware dependencies except at the motor servo level, would provide a flexible architecture for coordination. A system on these lines is being built to control an articulated hand and an arm. The four levels of control envisaged include a task decomposition level, a planning level, a scheduling level and a server level. The hand will carry both force and tactile sensors, feedback from these are used to provide adaptive control in grasping tasks. The processing of the sensory information is performed by independent processes, with analyzed information being sent to the relevant layer of the system. The manipulators are also controlled by individual processes. All process can open communications with an active process sending commands or data, or receiving them. We describe the scope of the system and the current setup plus future lines of development.
We develop a neural network formulation for multi-vehicle navigation on a two-dimensional surface. here. A time-linking map is generated for each individual vehicle using techniques similar to the known shortest path algorithms for an isolated vehicle. Neural networks are then applied to generate non-conflicting paths minimizing the time of travel.
An electronic neural network with feedback architecture, implemented in analog custom VLSI is described. Its application to problems of global optimization for dynamic assignment is discussed. The convergence properties of the neural network hardware are compared with computer simulation results. The neural network's ability to provide optimal or near optimal solutions within only a few neuron time constants, a speed enhancement of several orders of magnitude over conventional search methods, is demonstrated. The effect of noise on the circuit dynamics and the convergence behavior of the neural network hardware is also examined.
We are presenting a neural network optimization circuit for robust regression. The minimization of Least Square Error (LSE), L2 norm, cost functions is predominantly used for fitting functions to data points, but LSE approaches are highly sensitive to outlier points, or points that don't follow the trend. To alleviate this problem, we replace the L2 norm error function by an L1 norm error function which sums the absolute deviation of the errors and puts less weight on outlier points. An analog neural network optimization circuit is then developed to minimize the sum of the L1 norm error function. Simulation examples are presented on example data sets that compare the neural network solution with LSE solution .
If an autonomous robot system is to make effective use of a dextrous hand for assembly, it must be able to (1) reason about how objects are intended to fit together and design mating trajectories, (2) derive uncertainty constraints which accommodate perturbations from the nominal trajectories, (3) and determine how objects should be grasped to exert the forces expected during execution. This paper discusses a planning system which exploits the knowledge of symmetry incorporated in a group theory based reasoner to arrive at a nominal assembly plan. The plan is then refined by establishing bounds on the permissible uncertainty and required forces. A final stage plans force mediated interaction by incorporating multiple agents which enforce the wrench closure requirements and uncertainty constraints over the space of possible grasps.
In this paper, we discuss advanced architectures for distributed sensor networks. This includes the development of efficient algorithms for data-combination, noise removal, and information abstraction. Specifically, the we discuss the following issues in distributed signal processing: *A new signal processing architecture for networking spatially distributed sensors. *An improved method for integration of sensor information. *Fault-tolerance capabilities within the proposed signal processing architecture.
We consider the navigation of autonomous mobile machines, which are referred to as robots, through unknown terrains, i.e, terrains whose models are not a priori known. We deal with point-sized robots in two- and three-dimensional terrains and circular robots in two-dimensional terrains. The two-dimensional (three-dimensional) terrains are finite-sized and populated by an unknown, but, finite, number of simple polygonal (polyhedral) obstacles. The robot is equipped with a sensor system that detects all vertices and edges that are visible from its present location. In this context, we deal with two basic navigational problems. In the visit problem, the robot is required to visit a sequence of destination points, in a specified order, using the sensor system. In the terrain model acquisition problem, the robot is required to acquire the complete model of the terrain by exploring the terrain with the sensor. We present a framework that yields solutions to both the visit problem and the terrain model acquisition problem using a single approach. A point robot employs the restricted visibility graph and the visibility graph as the navigational course in two- and three-dimensional cases respectively. A circular robot employs the modified visibility graph. We present and analyze the algorithms to solve the visit problem and the terrain model acquisition problem based on the abovementioned structures.
Multilayer networks and recurrent neural networks have proved extremely successful in pattern recognition problems as well as in associative learning. In this paper an attempt is made to demonstrate that both types of networks, combined in arbitrary configurations, will find application in complex dynamical systems. Well known results in linear systems theory and their extensions to conventional adaptive control theory are used to suggest models for the identification and control of nonlinear dynamic systems. The use of neural networks in dynamical systems raises many theoretical questions, some of which are discussed in the paper.
At GTE Laboratories, we are advancing the theory of connectionist learning architectures for real-time control while exploring their relationships to animal learning models, applications in manufacturing quality control, and VLSI implementations. We seek connectionist-network architectures with improved convergence rate and scaling properties, as assessed on simulated and actual control problems. Our primary focus is on extensions to reinforcement learning. These include adaptive critics, feature/representation adaptation in multilayer networks, hybrid connectionist/conventional controllers, and modular networks for hierarchical control. We are also extending methods for system identification, or model learning, to include internal models learned using temporal-differences. We propose the integration of reinforcement and model learning based on their relationships to dynamic programming. We are working to resolve how connectionist systems should serve as a total systems concept or as tools in a larger architecture.
Consider the general problem of finding the inverse of dynamical systems. Given a dynamical system, it defines a causal relation between the inputs and the outputs. Both the inputs and the outputs are gnerally functions of time. The problem of dynamic inverse is to find an input for each given output such that the input produces an output that is in agreement with the given output in some sense. This problem of dynamic inverse may be formalized in the context of differential equations as the following. Given a system in the class of systems i=f(x,u) y = h(x) , x0 is given (1) where the state x is an n-vector, and the input u and output y are scalars. We want to find an input u(t) for each desired output r(t) such that the solution y(t) of Equation (1) with this input approximates r9t) in some sense.
We discuss here the basic elements of the new software system for large scale computation -MOVIE- (Metashell based Object oriented Visual Interactive Environment), recently designed and implemented at Caltech within the Caltech Concurrent Computation Program. From the research perspective, the goal of the MOVIE project is to create a simulation environment for modeling complex systems, with the focus on computational structures capable to adapt and act "intelligently", such as ensembles of image processing, early vision, neural network and Artificial Intelligence modules, integrated in the form of "neural robots". The high level MOVIE model, based on portable communication and computation protocol is suitable for large scale "intelligence engineering" by modeling such systems in distributed heterogeneous multicomputer environment and porting successful implementations to dedicated massively parallel hardware. From the software engineering point of view, the MOVIE model offers a platform for unifying elements of contemporary computing such as networking, windowing, parallelism, number crunching and symbolic processing. The basic idea, borrowed from Sun NeWS, is to use an appropriately extended PostScript as the unifying language. The MOVIE extension aims at promoting PostScript to a general purpose high level object oriented language with a high performance user expandable computational sector, fully compatible with the Adobe model for 2D graphics and the Sun X11 /NeWS model for windowing and multitasking.
The U.S. Bureau of Mines is currently investigating ways to combine the control capabilities of fuzzy logic with the learning capabilities of genetic algorithms. Fuzzy logic allows for the uncertainty inherent in most control problems to be incorporated into conventional expert systems. Although fuzzy logic based expert systems have been used successfully for controlling a number of physical systems, the selection of acceptable fuzzy membership functions has generally been a subjective decision. In this paper, high-performance fuzzy membership functions for a fuzzy logic controller that manipulates a mathematical model simulating the autonomous rendezvous of a spacecraft are learned using a genetic algorithm, a search technique based on the mechanics of natural genetics. The membership functions learned by the genetic algorithm provide for a more efficient fuzzy logic controller than membership functions selected by the authors for the rendezvous problem. Thus, genetic algorithms are potentially an effective and structured approach for learning fuzzy membership functions.
Simple Genetic Algorithms (SGA) have been shown to be useful tools for many function optimization problems. One present drawback of SGA is the time penalty involved in evaluating the fitness functions (performance indices) for large populations, generation after generation. This paper explores a small population approach (coined as Micro-Genetic Algorithms--μGA) with some very simple genetic parameters. It is shown that ,μGA implementation reaches the near-optimal region much earlier than the SGA implementation. The superior performance of the ,μGA in the presence of multimodality and their merits in solving non-stationary function optimization problems are demonstrated.