Sukhan Lee

A Gaussian potential function network with hierarchically self-organizing learning

Abstract This article presents a design principle of a neural network using Gaussian activation functions, referred to as a Gaussian Potential Function Network (GPFN), and explores the capability of a GPFN in learning a continuous input-output mapping from a given set of teaching patterns. The design principle is highlighted by a Hierarchically Self-Organizing Learning (HSOL) algorithm featuring the automatic recruitment of hidden units under the paradigm of hierarchical learning. A GPFN generates an arbitrary shape of a potential field over the domain of the input space, as an input-output mapping, by synthesizing a number of Gaussian potential functions provided by individual hidden units referred to as Gaussian Potential Function Units (GPFUs). The construction of a GPFN is carried out by the HSOL algorithm which incrementally recruits the minimum necessary number of GPFUs based on the control of the effective radii of individual GPFUs, and trains the locations (mean vectors) and shapes (variances) of individual Gaussian potential functions, as well as their summation weights, based on the Backpropagation algorithm. Simulations were conducted for the demonstration and evaluation of the GPFNs constructed based on the HSOL algorithm for several sets of teaching patterns.

Dual redundant arm configuration optimization with task-oriented dual arm manipulability

Cas de la manutention par deux bras cinematiquement redondants. Optimisation de la configuration des bras en jonction de la tâche a effectuer

Offline tracing and representation of signatures

An approach to the representation of signatures in an offline environment that makes a tracing of a signature in a manner similar to that of a human normally does, and incorporates the dynamic information of the tracing sequence into the representation of the signature, is presented. Tracing involves hierarchical decision-making for stroke identification and ordering based on a set of heuristic rules. These heuristic rules work as operators that transform a two-dimensional (2-D) spatial pattern into a one-dimensional (1-D) temporal pattern, thus making it possible to extract the dynamic features of a signature. Following the stroke sequence of a signature identified by the tracing, a multiresolution critical-point segmentation method is proposed to extract local feature points, at varying degrees of scale and coarseness, for representation. To make the representation translation-, rotation-, and scaling-invariant, a critical-point normalization method is introduced. Experimental results are shown. >

Computer control of space-borne teleoperators with sensory feedback

This paper presents the conceptual design, analysis, synthesis and software organization of an advanced teleoperator control system with sensory feedback. The design requirements for the system are discussed in detail and an implementation strategy is presented. The resulting system features maximum autonomy of the local hand controller and remote manipulator subsystems, along with kinematic and dynamic coordination between these subsystems. The final design emphasizes cooperation and interaction between the human operator and the computers in control of the sensor-based manipulator system. The hardware and software modules being used to implement the system at JPL are described.

Assembly planning based on geometric reasoning

Abstract This paper presents the application of geometric reasoning to the automatic construction of an assembly partial order from an attributed liaison graph representation of an assembly. The construction is based on the principle of assembly by disassembly and on the extraction of preferred subassemblies. On the basis of accessibility and manipulability criteria, the system first decomposes the given assembly into clusters of mutually inseparable parts. An abstract liaison graph is then generated wherein each cluster of mutually inseparable parts is represented as a supernode. A set of subassemblies is then generated by decomposing the abstract liaison graph into subgraphs, verifying the disassemblability of individual subgraphs, and applying the criteria for selecting preferred subassemblies to the disassemblable subgraphs. The recursive application of this process to the selected subassemblies results in a Hierarchical Partial-Order Graph (HPOG). A HPOG not only provides the precedence relations among assembly operations but also presents the temporal and spatial parallelism for implementing distributed and cooperative assembly. The system is organized under the “cooperative problem solving (CPS)” paradigm.

A Kalman filter approach for accurate 3-D motion estimation from a sequence of stereo images

Abstract This paper presents a Kalman filter approach for accurately estimating the 3-D position and orientation of a moving object from a sequence of stereo images. Emphasis is given to finding a solution for the following problem incurred by the use of a long sequence of images: the images taken from a longer distance suffer from a larger noise-to-signal ratio, which results in larger errors in 3-D reconstruction and, thereby, causes a serious degradation in motion estimation. To this end, we have derived a new set of discrete Kalman filter equations for motion estimation: (1) The measurement equation is obtained by analyzing the effect of white Gaussian noise in 2-D images on 3-D positional errors (instead of directly assigning Gaussian noise to 3-D feature points) and by incorporating an optimal 3-D reconstruction under the constraints of consistency satisfaction among 3-D feature points. (2) The state propagation equation, or the system dynamic equation, is formulated by describing the rotation between two consecutive 3-D object poses, based on quaternions and representing the error between the true rotation and the nominal rotation (obtained by 3-D reconstruction) in terms of the measurement noise in 2-D images. Furthermore, we can estimate object position from the estimation of object orientation in such a way that an object position can be directly computed once the estimation of an object orientation is obtained. Simulation results indicate that the Kalman filter equations derived in this paper represent an accurate model for 3-D motion estimation in spite of the first-order approximation used in the derivation. The accuracy of this model is demonstrated by the significant error reduction in the presence of large triangulation errors in a long sequence of images and by a shorter transition period for convergence to the true values.

Redundant arm kinematic control with recurrent loop

Abstract A new method for generating redundant arm inverse kinematic solutions based on the iterative update of joint vectors is presented. A novel neural network architecture with a recurrent loop is then formed based on the proposed method. In the proposed method, the pseudoinverse of the gradient of a Lyapunov function is defined in the joint space to update the joint vector toward a solution. This differs from the conventional methods based on the Jacobian pseudoinverse or Jacobian transpose. This paper establishes explicit convergence control schemes to achieve fast and stable convergence: 1) the convergence speed is enhanced by modifying the convergence dynamics based on terminal attractors, and 2) the convergence stability is ensured by the adaptive selection of update intervals based on the stability condition derived in this paper. The proposed neural network consists of a feedforward network and a feedback network forming a recurrent loop. The feedforward network is a multilayer network with hidden units having sinusoidal activation functions. As such, it computes accurately the forward kinematic solutions with simple training. The feedback network is derived from the feedforward network and computes joint vector updates. The proposed neural network has definite advantages over conventional neural networks for robot arm kinematic control, because it can not only handle redundant arm kinematic control, but can also provide an accurate computation of forward and inverse kinematic solutions with very simple training. The simulation results demonstrate that the proposed method is effective for the real-time kinematic control of a redundant arm as well as the real-time generation of collision-free joint trajectories.

Robot arm dynamic model reduction for control

General methods are described by which the mathematical complexities of explicit and exact state equations of robot arms can be reduced to a simplified and compact state equation representation without introducing significant errors into the robot arm dynamic model. The model reduction methods are based on homogeneous coordinates and on the Lagrangian algorithm for robot arm dynamics, and utilize matrix, vector and numeric analysis techniques. The derivation of differential vector representation of centripetal and Coriolis forces which has not yet been established in the literature is presented.

Automatic construction of assembly partial-order graphs

A system is presented which automatically constructs an assembly partial-order graph from an object-oriented model describing parts and their connections. The construction is based on the principle of assembly by disassembly and on assembly heuristics representing criteria for preferred subassemblies. The system first identifies disassemblable parts and subassemblies by reasoning geometric and physical constraints as well as resource requirements, and then assigns preference to the identified disassembly options based on assembly heuristics. The recursive application of this process to the selected disassemblies results in a hierarchical partial-order graph (HPOG). The HPOG not only specifies the required precedence in part assembly but also provides parallelism for implementing multiple-robot distributed and cooperative assembly. The software is organized under the cooperative-problem-solving paradigm.<<ETX>>

Collision-free path planning with neural networks

We present an efficient path planning approach which represents a path by a series of via points connected by elastic strings that are subject to displacement due to collisions with obstacles as well as constraints pertaining to the domain to which path planning is applied. Obstacle regions are represented by a potential field created by a multilayered neural network. A fast simulated annealing approach is used for local minima problems from the potential field. The automatic generation and removal of via points is incorporated in the path planning approach to ensure collision-free planning regardless of the complexity of the environment (e.g., convex, concave or complicated obstacle regions). Our path planning approach is flexible due to the automatic generation and removal of via points based on the complexity of the domain of the path planning process, efficient due to the use of the necessary via points for the path representation at all times, and massively parallel due to the parallel computation of motions of via points with only local information.