Milton Roberto Heinen

Applying Genetic Algorithms to Control Gait of Physically Based Simulated Robots

This paper describes our studies in the legged robots research area and the development of the LegGen System, that is used to automatically create and control stable gaits for legged robots into a physically based simulation environment. The parameters used to control the robot are optimized using genetic algorithms (GA). Comparisons between different robot models and fitness functions were accomplished, indicating how to compose a better multi-criterion fitness function to be used in the gait control of legged robots. The best gait control solution and the best robot model were selected in order to help us to build a real robot. The results also showed that it is possible to generate stable gaits using GA in an efficient manner.

An incremental probabilistic neural network for regression and reinforcement learning tasks

This paper presents a new probabilistic neural network model, called IPNN (for Incremental Probabilistic Neural Network), which is able to learn continuously probability distributions from data flows. The proposed model is inspired by the Spechts general regression neural network, but have several improvements which makes it more suitable to be used on-line in and robotic tasks. Moreover, IPNN is able to automatically define the network structure in an incremental way, with new units added whenever necessary to represent new training data. The performed experiments shows that IPNN is very useful in regression and reinforcement learning tasks.

IPNN: An Incremental Probabilistic Neural Network for Function Approximation and Regression Tasks

This paper presents a new probabilistic neural network model, called IPNN (for Incremental Probabilistic Neural Network), which is able to learn continuously probability distributions from data flows. The proposed model is inspired in the Spechts general regression neural network, but have several improvements which makes it more suitable to be used in on-line and robotic tasks. Moreover, IPNN is able to automatically define the network structure in an incremental and on-line way, with new units added whenever necessary to represent new training data. The experiments performed using the proposed model shows that IPNN is able to approximate continuous functions using few probabilistic units.

Applying Genetic Algorithms to Control Gait of Simulated Robots

This paper describes the LegGen simulator, used to automatically create and control stable gaits for legged robots into a physically based simulation environment. In our approach, the gait is defined using two different methods: a finite state machine based on robots leg joint angles sequences; and a recurrent neural network. The parameters for both methods are optimized using genetic algorithms. The model validation was performed by several experiments realized with a robot simulated using the ODE physical simulation engine. The results showed that it is possible to generate stable gaits using genetic algorithms in an efficient manner, using these two different methods.

Feature-Based Mapping Using Incremental Gaussian Mixture Models

This paper proposes a new algorithm for feature-based environment mapping where the environment is represented using multivariate Gaussian mixture models. This algorithm, which can be used either with sonar or laser range data, is able to create and maintain environment maps in real time using few memory requirements. Moreover, it does not assume that the environment is composed by linear structures and allows computing the occupancy probabilities of any map position very fast and without introducing discretization errors. The experiments performed with the proposed model prototype show that it is able to build accurate environment representations using real data provided by a mobile robot.

Evaluation of visual attention models under 2D similarity transformations

The computational models of visual attention, originally proposed as cognitive models of human attention, nowadays are being used as front-ends to some robotic vision systems, like automatic object recognition and landmark detection. However, these kinds of applications have different requirements from those originally proposed. More specifically, a robotic vision system must be relatively insensitive to 2D similarity transformations of the image, as in-plane translations, rotations, reflections, and scales. In this paper several experiments with two visual attention models publicly available are described. The results show that the best known model, called NVT, is extremely sensitive to these 2D similarity transformations. Therefore, a new visual attention model, called NLOOK, is proposed and validated with the same invariance criteria, and the results show that NLOOK is less sensitive to these kind of transformations than the other two models. Besides, NLOOK can select better fixations according to a redundancy criterion. Thus, the proposed model is an excellent tool to be used in robot vision systems.

Incremental feature-based mapping from sonar data using Gaussian mixture models

This paper proposes a new algorithm for feature-based environment mapping where the environment is represented using multivariate Gaussian mixture models. This algorithm, which can be used either with sonar or laser range data, is able to create and maintain environment maps in real time using few memory requirements. Moreover, it does not assume that the environment is composed by linear structures and allows computing the occupancy probabilities of any map position very fast and without introducing discretization errors. The experiments performed using sonar data show that it is able to build accurate environment representations using noisy data provided by a mobile robot.

Recursive Incremental Gaussian Mixture Network For Spatio-Temporal Pattern Processing

This work introduces a novel neural network algorithm for online spatio-temporal pattern processing, called Echo State Incremental Gaussian Mixture Network (ESIGMN). The proposed algorithm is a hybrid of two stateof-the-art algorithms: the Echo State Network (ESN), used for spatio-temporal pattern processing, and the Incremental Gaussian Mixture Network (IGMN), applied to aggressive learning in online tasks. The algorithm is compared against the conventional ESN in order to highlight the advantages of the IGMN approach as a supervised output layer. Resumo. Este trabalho introduz um novo algoritmo de redes neurais para processamento online de padroes espaco-temporais, chamado Echo State Incremental Gaussian Mixture Network (ESIGMN). O algoritmo proposto e um hibrido de dois algoritmos estado-da-arte: a Echo State Network (ESN), usada para processamento de padroes espaco-temporais, e a Incremental Gaussian Mixture Network (IGMN), aplicada ao aprendizado agressivo em tarefas online. O algoritmo e comparado com a ESN convencional a fim de destacar as vantagens da abordagem IGMN como camada supervisionada de saida.

Visual Selective Attention Model for Robot Vision

This paper describes a model of visual selective attention, called NLOOK, proposed to be used in computational and robotic vision systems. This model first decomposes the visual input in a set of topographic feature maps which encode intensity, orientation, color and movement. All feature maps feed into a master ldquosaliency maprdquo, which topographically codifies for local conspicuity over the entire visual scene, and a winner-take-all neural network with an inhibition of return mechanism that selects the most salient points of the map in decreasing order. The obtained results demonstrate that the proposed model is suitable for robotic vision systems.

Aprendizado e Controle de Robôs Móveis Autônomos Utilizando Atenção Visual

Este artigo descreve um modelo de aprendizado por reforco capaz de aprender tarefas de controle complexas utilizando acoes e estados continuos. Este modelo, que e baseado no ator-critico continuo, utiliza redes de funcoes de base radial normalizadas para aprender o valor dos estados e das acoes, sendo capaz de configurar a estrutura destas redes de forma automatica durante o aprendizado. Alem disso, um mecanismo de atencao visual seletiva e utilizado para perceber o ambiente e os estados. Para a validacao do modelo proposto, foi utilizada uma tarefa relativamente complexa para os algoritmos de aprendizado por reforco: conduzir uma bola ate o gol em um ambiente de futebol de robos simulado. Os experimentos realizados demonstram que o modelo proposto e capaz realizar a tarefa em questao com bastante sucesso utilizando somente informacoes visuais.