Catálogo de publicaciones - libros

Presented is a spiking neural network architecture of human language instruction recognition and robot control. The network is based on a model of a leaky Integrate-And-Fire (lIAF) spiking neurone with Active Dendrites and Dynamic Synapses (ADDS) [1,2,3]. The architecture contains several main modules associating information across different modalities: an auditory system recognising single spoken words, a visual system recognising objects of different colour and shape, motor control system for navigation and motor control and a working memory. The main focus of this presentation is the working memory module whose function is sequential processing of word from a language instruction, task and goal representation and cross-modal association of objects and actions. We test the model with a robot whose goal is to recognise and execute language instructions. The work demonstrates the potential of spiking neurons for processing spatio-temporal patterns and the experiments present spiking neural networks as a paradigm which can be applied for modelling sequence detectors at word level for robot instructions.

We present a virtual reality platform for developing and evaluating embodied models of cognitive development. The platform facilitates structuring of the learning agent, of its visual environment, and of other virtual characters that interact with the learning agent. It allows us to systematically study the role of the visual and social environment for the development of particular cognitive skills in a controlled fashion. We describe how it is currently being used for constructing an embodied model of the emergence of gaze following in infant-caregiver interactions and discuss the relative benefits of virtual vs. robotic modeling approaches.

In order to bootstrap shared communication systems, robots must have a non-verbal way to influence the attention of one another. This chapter presents an experiment in which a robot learns to interpret pointing gestures of another robot. We show that simple feature-based neural learning techniques permit reliably to discriminate between left and right pointing gestures. This is a first step towards more complex attention coordination behaviour. We discuss the results of this experiment in relation to possible developmental scenarios about how children learn to interpret pointing gestures.

Function approximators are commonly in use for reinforcement learning systems to cover the untrained situations, when dealing with large problems. By doing so however, theoretical proves on the convergence criteria are still lacking, and practical researches have both positive and negative results. In a recent work [3] with neural networks, the authors reported that the final results did not reach the quality of a -table in which no approximation ability was used. In this paper, we continue this research with grid based function approximators. In addition, we consider the required number of state transitions and apply ideas from the field of active learning to reduce this number. We expect the learning process of a similar problem in a real world system to be significantly shorter because state transitions, which represent an object’s actual movements, require much more time than basic computational processes.

A biologically inspired computational model of rodent repre-sentation–based (locale) navigation is presented. The model combines visual input in the form of realistic two dimensional grey-scale images and odometer signals to drive the firing of simulated place and head direction cells via Hebbian synapses. The space representation is built incrementally and on-line without any prior information about the environment and consists of a large population of location-sensitive units (place cells) with overlapping receptive fields. Goal navigation is performed using reinforcement learning in continuous state and action spaces, where the state space is represented by population activity of the place cells. The model is able to reproduce a number of behavioral and neuro-physiological data on rodents. Performance of the model was tested on both simulated and real mobile Khepera robots in a set of behavioral tasks and is comparable to the performance of animals in similar tasks.

To have agents autonomously model a complex environment, it is desirable to use distributed representations that lend themselves to neural learning. Yet developing and executing plans acting on the environment calls for abstract, localist representations of events, objects and categories. To combine these requirements, a formalism that can express neural networks, action sequences and symbolic abstractions with the same means may be considered advantageous. We are currently exploring the use of compositional hierarchies that we treat both as and as localist representations for plans and control structures. These hierarchies are implemented using and used in the control of agents situated in a complex simulated environment.

In this paper we present our bio-mimetic artificial arm and the simulation results on its low level control system. In accordance with the general view of the Biorobotics field we try to replicate the structure and the functionalities of the natural limb. The control system is organized in a hierarchical way, the low level control reproduces the human spinal reflexes and the high level control the circuits present in the cerebral motor cortex and the cerebellum. Simulation show how the system controls the single joint position reducing the stiffness during the movement.

This paper presents a human-like control of an innovative biped robot. The robot presents a total of twelve degrees of freedom; each joint resemble the functionalities of the human articulation and is moved by tendons connected with an elastic actuator located in the robot’s pelvis. We implemented and tested an innovative control architecture (called elastic-reactive control) that permits to vary the joint stiffness in real time maintaining a simple position-control paradigm. The controller is able to estimate the external load measuring the spring deflection and demonstrated to be particularly robust respect to system uncertainties, such as inertia value changes. Comparing the resulting control law with existing models we found several similarities with the Equilibrium Point Theory.

Biologically inspired robotics offers the promise of future autonomous devices that can perform significant tasks while coping with noisy, real-world environments. In order to survive for long periods we believe a developmental approach to learning is required and we are investigating the design of such systems inspired by results from developmental psychology. Developmental learning takes place in the context of an epigenetic framework that allows environmental and internal constraints to shape increasing competence and the gradual consolidation of control, coordination and skill. In this paper we describe the use of novelty and habituation as the motivation mechanism for a sensory-motor learning process. In our system, a biologically plausible habituation model is utilized and the effect of parameters such as habituation rate and recovery rate on the learning/development process is studied. We concentrate on the very early stages of development in this work. The learning process is based on a topological mapping structure which has several attractive features for sensory-motor learning. The motivation model was implemented and tested through a series of experiments on a working robot system with proprioceptive and contact sensing. Stimulated by novelty, the robot explored its egocentric space and learned to coordinate motor acts with sensory feedback. Experimental results and analysis are given for different parameter configurations, proprioceptive encoding schemes, and stimulus habituation schedules.

This chapter explores modular learning in artificial neural networks for intelligent robotics. Mainly inspired from neurobiological aspects, the modularity concept can be used to design artificial neural networks. The main theme of this chapter is to explore the organization, the complexity and the learning of modular artificial neural networks. A robust modular neural architecture is then developed for the position/orientation control of a robot manipulator with visual feedback. Simulations prove that the modular learning enhances the artificial neural networks capabilities to learn and approximate complex problems. The proposed bidirectional modular learning architecture avoids the neural networks well-known limitations. Simulation results on a 7 degrees of freedom robot-vision system are reported to show the performances of the modular approach to learn a high-dimensional nonlinear problem. Modular learning is thus an appropriate solution to robot learning complexity due to limitations on the amount of available training data, the real-time constraint, and the real-world environment.