(Return to ICNN97 Homepage) (Return to ICNN'97 Agenda)
ICNN'97 FINAL ABSTRACTS
RV: ROBOTICS AND MACHINE VISION
ICNN97 Robotics and Machine Vision Session: RV1A Paper Number: 207 Oral
Modified 3-D Hopfield neural network for gesture recognition
Chung-Lin Huang and Wen-Yi Huang
Keywords: 3-D Hopfield neural network gesture recognition Hausdorff distance measure
Abstract:
_____
ICNN97 Robotics and Machine Vision Session: RV1B Paper Number: 244 Oral
Range image segmentation using an oscillatory network
Xiuwen Liu and DeLiang Wang
Keywords: Range image segmentation oscillatory network legion network
Abstract:
We use a locally excitatory globally inhibitory oscillator network (LEGION) as a framework for range image segmentation. Each oscillator in the LEGION net work has excitatory lateral connections to the oscillators in its neighborhood as well as a connection with a global inhibitor. The lateral connection between two oscillators is established based on the similarity between their feature vectors which consist of the surface normal and curvature at the corresponding pixel locations. The emergent behav ior of the LEGION network gives rise to the segmentation result. Unlike other methods, our scheme needs no assumption about the underlying structures in image data and no prior knowledge regarding the number of regions. Experimental results for real range images are presented.
_____
ICNN97 Robotics and Machine Vision Session: RV1C Paper Number: 532 Oral
An improved neural network for segmentating objects' boundaries in real images
Wee Kheng Leow and Seet Chong Lua
Keywords: Boundaries identification Boundary contour system Object segmentation
Abstract:
An important task in object recognition is to first identify the boundaries of the objects in the input image. Several neural networks have been proposed to perform edge detection and boundary segmentation. Among them, Grossberg and Mingolla's Boundary Contour System (BCS) seems promising because it is able to complete missing object boundaries. Although BCS has been shown to work well on synthetic and silhouette images, we found that it has some shortcomings when applied to real images. This paper presents an improved version of BCS for handling the shortcomings.
_____
ICNN97 Robotics and Machine Vision Session: RV1D Paper Number: 180 Oral
Modelling active contours using neural networks isomorphic to boundaries
Y.V. Venkatesh and N. Rishikesh
Keywords: active contours deformable templates deformation of patterns isomorphism
Abstract:
We propose a new technique, based on self-organization, for localizing salient contours in an image, with applications to, for instance, object and character recognition, stereopsis and motion tracking. A neuronal network which is isomorphic to the template/initial contour is created. This network acts as an active contour, which, using self-organization, undergoes deformation in an attempt to cling on to the nearest salient contour in the test image. The application areas of the model proposed are similar to those of the 'snake' model of the literature. But the proposed model is different from the 'snake' model in both the underlying mathematics and implementation. The new technique is illustrated with some examples.
_____
ICNN97 Robotics and Machine Vision Session: RV1E Paper Number: 277 Oral
Evolutionary computation for figure-ground separation
Suchendra M. Bhandarkar and Xia Zeng
Keywords: Evolutionary computation figure-ground separation energy minimization
Abstract:
The problem of figure-ground separation is modeled as one of energy minimization using the Ising system model from quantum physics. The Ising system model for the figure-ground separation problem makes explicit the definition of shape in terms of attributes such as cocircularity, smoothness, proximity and contrast and is based on the formulation of an energy function that incorporates pairwise interactions between local image features in the form of edgels. The paper explores a class of stochastic optimization techniques based on evolutionary algorithms in the context of figure-ground separation using the Ising system model. Experimental results on synthetic edgel maps and edgel maps derived from gray scale images are presented.
_____
ICNN97 Robotics and Machine Vision Session: RV1F Paper Number: 517 Oral
Pixel Based 3D object recognition with bidirectional associative memory
I. Elsen, K. F. Kraiss, D. Krumbiegel
Keywords: 3D object recognition Bidirectional assocoative memory
Abstract:
This paper addresses the pixel based recognition of 3D--objects with bidirectional associative memories. Computational power and memory requirements for this approach are identified and compared to the performance of current computer architectures by benchmarking different processors. It is shown, that the performance of special purpose hardware, like neurocomputers, is between one and two orders of magnitude higher than the performance of mainstream hardware. On the other hand, the calculation of small neural networks is performed more efficiently on mainstream processors. Based on these results a novel concept is developed, which is tailored for the efficient calculation of bidirectional associative memories. The computational efficiency is further enhanced by the application of algorithms and storage techniques which are matched to characteristics of the application at hand.
_____
ICNN97 Robotics and Machine Vision Session: RV2A Paper Number: 547 Oral
Trajectory control of robotic manipulators using chaotic neural networks
Sang-Hee Kim, Cahng-Wha Jang, Chang-Hyun Chai and Han-Go Choi
Keywords: trajectory control robotic manipulators chaotic neural networks
Abstract:
This paper investigates the direct adaptive control of nonlinear systems using chaotic neural networks. Since the structure of a chaotic neural networks contain self and internal feedback loops in each layer, chaotic neural networks can show the robust characteristics for controlling highly nonlinear dynamics like robotic manipulators. This paper presents modified chaotic neural networks with the backpropagation learning algorithm. To evaluate the performance of the proposed neural networks, we simulate the trajectory control of the three-axis PUMA robot with direct adaptive control strategies. The structure of the robot controller consists of the PD controller and chaotic neural networks controller in parallel. Simulation results showed the superior performance on convergence and final error comparing with recurrent neural networks. Chaotic neural networks also reduce the number of nodes and computation time.
_____
ICNN97 Robotics and Machine Vision Session: RV2B Paper Number: 317 Oral
A partially recurrent gating network approach to learning action selection by reinforcement
R.M. Rylatt, C.A. Czarnecki and T.W. Routen
Keywords: partially recurrent gating network action selection reinforcement
Abstract:
We describe a nueral network approach to the problem of reactive navigation, using a simulated mobile robot. Specifically, it is shown that complementary reinforcement backpropagation learning can be a means for modular networks to acquire different navigation related skills concurrently. Further, it is demonstrated that a partially recurrent net can function as a gating network to co-ordinate the reinforcement learning across time steps. In effect, the recurrent gating network performs action selection by choosing developing experts to make control decisions in the context of previous actions in the temorally extended domain.
Keywords: action selection, complementary reinforcement backpropagation, gating networks, reactive navigation.
_____
ICNN97 Robotics and Machine Vision Session: RV2C Paper Number: 538 Oral
Adaptive learning with the growing competitive linear local mapping network for robotic hand-eye coordination
Andrei Cimponeriu and Julien Gresser
Keywords: Linear local mapping Learning Kohonen Map
Abstract:
Traditionally, Linear Local Mapping Networks learn the entire workspace, and the neurons are placed according to the Kohonen map or its variant, the "neural gas". In this paper a new neural network is introduced, which allocates neurons adaptively following the current trajectory, according to an error criterion. The resulting network has a small number of neurons and is thus very efficient. It also learns very quickly: employing an active learning and the RLS algorithm, just one pass is sufficient for our algorithm to acquire the Jacobians that are needed to perform a given positioning of robot's gripper on the target. Also, an on-line adaptation of the Jacobians is proposed.
_____
ICNN97 Robotics and Machine Vision Session: RV2D Paper Number: 409 Oral
Self-organizing geometric certainty maps: A compact and multifunctional approach to map building, place recognition and motion planning
J. A. Janet, M. W. White, J. C. Sutton, III and W. E. Snyder
Keywords: Sel-organizing Kohonen Neural net Hyperellipsoidal clustering self localization Kolmogorov-Smirnov Test
Abstract:
In this paper we show how a self-organizing Kohonen neural network using hyperellipsoid clustering (HEC) can build maps from actual sonar data. With the HEC algorithm we can use the Mahalanobis distance to learn elongated shapes (typical of sonar data) and obtain a stochastic measurement of data-node association. Hence, the HEC Kohonen can be used to build topographical maps and to recognize its own topographical cues for self-localization. The number of nodes can also be regulated in a self-organizing manner by measuring how well a node models the statistical properties of its associated data. This measurement determines whether a node should be divided (mitosis) or pruned completely. Because fewer nodes are needed for an HEC Kohonen than for a Kohonen that uses only Euclidean distance, the data size is smaller for the HEC Kohonen. Relative to grid-based approaches, the savings in data size is even more profound. By incorporating principal component analysis (PCA), the HEC Kohonen can handle problems with several dimensions (3-D, time-series, etc.). The HEC Kohonen is also multifunctional in that it accommodates compact geometric motion planning and self-referencing algorithms. It can also be generalized to solve other pattern recognition problems.
_____
ICNN97 Robotics and Machine Vision Session: RV2E Paper Number: 369 Oral
Fuzzy logic and neural network based adaptive controller design
Ya-chen Hsu, Guanrong Chen and Heider A. Malki
Keywords: Fuzzy logic neural network adaptive controller design
Abstract:
An adaptive control algorithm based on the sliding mode principle, equipped with fuzzy logic to handle system modeling uncertainties, is developed in this paper. Along with a neural-network learning scheme that enhances the adaptive control capability, this controller performs satisfactory tracking control for a large class of robot models that contain significant but unknown friction and disturbances. Both mathematical analysis and computer simulation are enclosed for demonstration.
_____
ICNN97 Robotics and Machine Vision Session: RV2F Paper Number: 540 Oral
recurrent neural network with self-adaptive GAs for biped locomotion robot
Yoichiro Komata
Keywords: recurrent neural network self-adaptive GAs biped locomotion robot
Abstract:
In this paper, we propose a generation method of a stable motion of a biped locomotion robot. We apply the proposed method to eight force sensors at the soles of the biped locomotion robot. Zero Moment Point (ZMP) is a well known as the index of stability in walking robots. ZMP is determined by the configuration of the robots. However, there are many configurations against the ZMP. Because of that, when we use ZMP as stabilization index, we must select the best among many stability configurations. Then it is a problem that which configuration is selcted. In this paper, the problem can be solved with recurrent neural network. We calculate the position of ZMP and the joints and the angles that should be actuated can be determined by recurrent neural network without ZMP moving out from the supporting area of sole.
We employ the recurrent neural network with self-adaptive GAs for learning capability. Further, we build a biped locomotion robot in trial, which has 13 joints and verified that the calculated stability motion trajectory can be successfully applied to the practical biped locomotion. In this paper, we propose a way of training of reccurrent neural network for biped locomotion robot.
_____
ICNN97 Robotics and Machine Vision Session: RVP1 Paper Number: 606 Poster
Neural force/position control in cartesian space for a 6 DOF industrial robot: concept and first results
R. Maas, V. Zahn and R. Eckmiller
Keywords: force/position control cartesian space industrial robot
Abstract:
A novel concept of neural force/position control in Cartesian space (NFC) was developed and applied. The NFC concept for a 6 DOF industrial robot with a 6 DOF sensor (3 * forces, 3 * torques) is based on a cycle time of just 2 msec. NFC features include: 1) Sensor data and trajectory input processing in Cartesian space; 2) Learned mapping operations for force, kinematics, and dynamics with neural networks; 3) Singularity robustness in the entire workspace; 4) Automatic adjustment of desired trajectories to kinematic and dynamic constraints. NFC allows the control of 6 DOF robots in defined contact with moving stiff objects and surfaces. This requires the dynamic handling of coupled force and position vectors. Results from simulations and real time experiments with a 6 joint manipulator (Siemens manutec r2) will be discussed.
_____
ICNN97 Robotics and Machine Vision Session: RVP1 Paper Number: 402 Poster
Robust telemanipulator control using a partitioned neural network architecture
Spyros G. Tzafestas, Platon A. Prokopiou and Costas S. Tzafestas
Keywords: Robust telemanipulator control partitioned neural network architecture control problem
Abstract:
In this paper the control problem of telemanipulators is considered under the condition that they are subject to modeling and other uncertainties of considerable levels. The design is based on the S. Lee and H. S. Lee teleoperator control scheme, which is modified so as to be able to compensate the uncertainties,and is implemented using a partitioned multilayer perceptron neural network. Several subnetworks are used each one identifying a term of the manipulator's dynamic model. A new learning algorithm is proposed which distributes the learning error to each subnetwork and enables online training. Several simulation results are provided, which show the robustness ability by the partitioned neurocontroller, and compare it with the results obtained through sliding mode control.
_____
ICNN97 Robotics and Machine Vision Session: RVP1 Paper Number: 375 Poster
Model-based path planning and tracking using neural networks for a robot manipulator
Sangbong Park and Cheol Hoon Park
Keywords: path planning path tracking neural networks robot manipulator
Abstract:
_____
ICNN97 Robotics and Machine Vision Session: RVP1 Paper Number: 621 Poster
Reinforcement learning when visual sensory signals are directly given as inputs
Katsunari Shibata and Yoichi Okabe
Keywords: Reinforcement learning visual sensory signals neural network
Abstract:
It is shown that a neural-network based learning system, which obtains visual signals as inputs directly from visual sensors, can modify its outputs by reinforcement learning. Even if each visual cell covered only a local receptive field, the learning system could integrate these visual signals and obtain a smooth evaluation function. It also represented the spatial information smoothly in the hidden layer through the learning, and the area of the state which seemed important for the system was magnified in the hidden neurons' space. The learning is so adaptive that when different motion characteristic was employed in the system, the representation became different from the previous one, even if the environment is the same.
_____
ICNN97 Robotics and Machine Vision Session: RVP1 Paper Number: 274 Poster
neural control of artificial six-legged terrestrial locomotion on rough terrain
Alejandro Ramirez-Serrano
Keywords: neural control artificial six-legged terrestrial locomotion knowledge-based neural network
Abstract:
This paper describes neural control of walking and leg locomotion on rough uneven terrains of a six-legged terrestrial robot. Central and leg control using neural networks is described. Rhythmic walking patterns are achieved by a central control using a knowledge-based neural network. On the other hand, local leg control is performed by an Adaline type neural network. Both networks were design based on biological principles. In order to produce movements similar to natural behavior, both networks receive feedback from leg sensors. Simulation outputs are presented as digital diagrams.
_____
ICNN97 Robotics and Machine Vision Session: RVP1 Paper Number: 329 Poster
Reinforcement control via action dependent heuristic dynamic programming
K. Wendy Tang and Govardhan Srikant
Keywords: Reinforcement control heuristic dynamic programming adaptive critic
Abstract:
Heuristic Dynamic Programming (HDP) is the simplest kind of Adaptive Critic which is a powerful form of reinforcement control. It can be used to maximize or minimize any utility function, such as total energy or trajectory error, of a system over time in a noisy environment. Unlike supervised learning, adaptive critic design does not require the desired control signals be known. Instead, feedback is obtained based on a critic network which learns the relationship between a set of control signals and the corresponding strategic utility function. It is an approximation of dynamic programming. Action Dependent Heuristic Dynamic Programing (ADHDP) system involves two subnetworks, the Action network and the Critic network. Each of these networks includes a feedforward and a feedback component. A flow chart for the interaction of these components is included. To further illustrate the algorithm, we use ADHDP for the control of a simple, 2-D planar robot.
_____
ICNN97 Robotics and Machine Vision Session: RVP1 Paper Number: 665 Poster
Fuzzy Q-learning for autonomous robot systems
Il Hong Suh, Jae-Hyun Kim and Frank Chung-Hoon Rhee
Keywords: Fuzzy Q-learning autonomous robot systems fuzzy interpolation
Abstract:
It is desirable for autonomous robot syustems to posses the ability to behave in a smooth and continous fashion when interacting with an unknown environment. Since Q-learning is normally used for optimizing a series of descrete actions, it may be undesirable when applied to a real environment which involves continuous states and actions.
In this paper, we propose a new method of Q-learning that incorporates a fuzzy interpolation technique which is used to approximate a continuous state. Our lerning method can estimate a current state by its neighboring states and has the ability to learn its actions similar to that of Q-learning. thus, our method can enable robots to react smoothly in a real environment. Simulation results involving an autonomous robot are given to show the validity of our method.
_____
ICNN97 Robotics and Machine Vision Session: RVP1 Paper Number: 640 Poster
Artificial neural networks and the simulation of human movements in CAD environments
M. Costa, P. Crispino, A. Hanomolo and E. Pasero
Keywords: Artificial neural networks simulation of human movements CAD
Abstract:
Keywords: Human-Like Movements, Ergonomics, 3D CAD
A powerful goal for 3D CAD tools is today the simulation of working environments to find the ergonomic parameters which relate people to objects in the space. Mathematical models of human movements are complex to define and hard to solve. Artificial Neural Systems (ANS) could be an interesting approach to the problem: why not to use artificial neurons to simulate actual neurons? This work deals with a first approach to a major project. We used ANS to simulate the movement of a human arm in a CAD environment.
_____
ICNN97 Robotics and Machine Vision Session: RVP1 Paper Number: 511 Poster
Image estimation and compensation based on correspondence Technique
Y. H. Kuo, M. F. Horng
Keywords: motion estimation motion compensation Correspondence Hopfield network
Abstract:
An algorithm based on correspondence technique for image motion estimation and compensation is proposed. In conventional methods, the motion estimation and compensation are frequently solved with a block-matching based strategy, making it computationally intensive. In fact, the correspondence between feature points in successive image sequence is an useful information to estimate the object motion in picture. From the correspondence found by a Hopfield neural network, the vectors, called displacement vectors, are constructed rapidly, and they will characterize the object motion process. By means of the massive parallel processing-power of neural network and the proposed computational model, an efficient and accurate solution can be obtained.
_____
ICNN97 Robotics and Machine Vision Session: RVP1 Paper Number: 130 Poster
New neural network for adaptive control of robot manipulators
Sahin Yilidirim
Keywords: neural network adaptive control robot manipulator
Abstract:
_____
ICNN97 Robotics and Machine Vision Session: RVP1 Paper Number: 122 Poster
Neurofuzzy agents and neurofuzzy laws for autonomous machine learning and control
Wen-Ran Zhang
Keywords: Neurofuzzy agents neurofuzzy laws autonomous machine learning
Abstract:
Real world autonomous agents exhibit adaptive, incremental, exploratory, and sometimes explosive learning behaviors. Learning in neurofuzzy control, however, is often referred to as global training with a large set of random examples and with a very low learning rate. This type of controller does not show exploratory learning behaviors. An agent-oriented approach to neurofuzzy control is introduced and illustrated in folding-legged robot locomotion and gymnastics; necessary and sufficient conditions are established for agent- oriented neurofuzzy discovery; and a theory of coordinated multiagent neurofuzzy control is analytically formulated. The analytical features bridge a gap between linear control, neurofuzzy control, adaptive learning, and exploratory learning.
_____
ICNN97 Robotics and Machine Vision Session: RVP1 Paper Number: 624 Poster
A neural network framework for low-level representation and processing in computer vision
Richard Lepage and Denis Poussart
Keywords: neural network low-level representation and processing computer vision
Abstract:
A goal of computer vision is the construction of scene descriptions based on information extracted from one or more 2D images. A reconstruction strategy based on a three-level representational framework is proposed. The first representational level, the Primal Sketch, makes explicit physical characteristics of the scene through detection of illuminance changes and their geometrical distribution and organization. Physical characteristics appear at several spatial scales and a multiresolution analysis helps in eliminating spurious edges. The second representational level, the raw 2.5D Sketch, makes explicit the orientation and rough depth at edge location of the visible surfaces. A multiresolution neural network stereo algorithm is designed to compute the disparity at each edge location and at all the resolution levels. Matching is facilitated by a hierarchical focussing mechanism. The third representation level, the full 2.5D Sketch, makes explicit the orientation and depth estimate at all the visible surface coordinates. Depth information between the edges is computed with a local shape-from-shading algorithm. A constraint satisfaction network fuses stereo and shading data.
_____
ICNN97 Robotics and Machine Vision Session: RVP1 Paper Number: 59 Poster
A neural network for the trajectory control of robotic manipulators with uncertainties
Boo Hee Nam, Sang Jae Lee and Seok Won Lee
Keywords: nonlinearity uncertainty computed torque method neural network
Abstract:
We propose a neural network to compensate for the structured and unstructured uncertainties in the robot model with the computed torque method. The neural network is used not to learn the inverse dynamic model but to compensate for the uncertainties of robotic manipulators.
When training the neural network, we use the teaching signals present in the proposed control scheme, whose control structure is simpler than that proposed by Ishiguro et al., whose teaching signals come from the robot model.