Current machine learning techniques proposed to automatically discover a robot's kinematics usually rely on a priori information about the robot's structure, sensor properties or end-effector position. This paper proposes a method to estimate a certain aspect of the forward kinematics model with no such information. An internal representation of the end-effector configuration is generated from unstructured proprioceptive and exteroceptive data flow under very limited assumptions. A mapping from the proprioceptive space to this representational space can then be used to control the robot.

A new method for self-supervised sensorimotor learning of sound source localization is presented, that allows a simulated listener to learn online an auditorimotor map from the sensorimotor experience provided by an auditory evoked behavior. The map represents the auditory space and is used to estimate the azimuthal direction of sound sources. The learning mainly consists in non-linear dimensionality reduction of sensorimotor data. Our results show that an auditorimotor map can be learned, both from real and simulated data, and that the online learning leads to accurate estimations of azimuthal sources direction.

Developmental Robotics offers a new approach to numerous AI features that are often taken as granted. Traditionally, perception is supposed to be an inherent capacity of the agent. Moreover, it largely relies on models built by the system's designer. A new approach is to consider perception as an experimentally acquired ability that is learned exclusively through the analysis of the agent's sensorimotor flow. Previous works, based on H.Poincaré's intuitions and the sensorimotor contingencies theory, allow a simulated agent to extract the dimension of geometrical space in which it is immersed without any a priori knowledge. Those results are limited to infinitesimal movement's amplitude of the system. In this paper, a non-linear dimension estimation method is proposed to push back this limitation.

Perception and action are fundamental tasks for autonomous robots. Traditionnally, they rely on theoretical models built by the system's designer. But, is a naive agent able to learn by itself the structure of its interaction with the environment without any a priori information? This knowledge should be extracted through the analysis of the only information it has access to: its high-dimensional sensorimotor flow. Recent works, based on the sensorimotor contingencies theory, allow a simulated agent to extract the geometrical space dimensionality without any model of itself nor of the environment. In this paper, these results are validated using a more sophisticated auditive modality. The question of multimodality fusion is then addressed by fitting up the agent with vision. Finally, preliminary experimental results on a real robotic platform are presented.

In this paper, we propose an extension of a self-organizing map called self-organizing multilayer perceptron (SOMLP) whose purpose is to achieve quantization of spaces of functions. Based on the use of multilayer perceptron networks, SOMLP comprises the unsupervised as well as supervised learning algorithms. We demonstrate that it is possible to use the commonly used vector quantization algorithms (LVQ algorithms) to build new algorithms called functional quantization algorithms (LFQ algorithms). The SOMLP can be used to model nonlinear and/or nonstationary complex dynamic processes, such as speech signals. While most of the functional data analysis (FDA) research is based on B-spline or similar univariate functions, the SOMLP algorithm allows quantization of function with high dimensional input space. As a consequence, classical FDA methods can be outperformed by increasing the dimensionality of the input space of the functions under analysis. Experiments on artificial and real world examples are presented which illustrate the potential of this approach.

First task robots have to realise is sensing and acting in the environment. Can a robot learn the way it is able to sense and act in the world without any hardwired notions? Is it able to learn it from the only data he has access to, that is high-dimension sensory inputs and motor outputs? This paper presents experimental results obtained on a simulated human listener using a bio-inspired model of the cochlea and real records from human related transfer functions (HRTF). These results show that a naive system that interacts with its environment without knowing the laws governing these interactions can discover information about dimensionality of space. Moreover, the laws determining the sensations of the system as a function of the state of the system and the environment, called the ¿sensorimotor law¿, are not simplified as usually in simulations. They are bio-realistic as they are determined by the HRTF recorded on human beings.

The project ASAROME (autonomous sailing robot for oceanographic measurements) is working on a small autonomous sailboat in order to make measurements and observations in the marine environment for long periods. In this project, perception plays an important role by giving an estimate of the speed of surface winds, the state of the sea surface and the rate of precipitation in wet weather. In this paper, the unknown signals are first encoded with different codes (ERB, MFCC, LPC, LPCC). Then the coded signals are modeled by two different methods of classification: predictive and k-nearest neighbor. The final part of the system uses local and global decision to recognize the class of the unknown signal. Experiments are conducted to compare the results obtained by different encodings. Our results show that MFCC does not represent the ideal approach for the recognition of underwater audio signals, but LPCC seems to be a better candidate.

In mobile robotics applications, pattern and object recognition are mainly achieved relying only on vision. Several other perceptual modalities are also available such as, touch, hearing or vestibular proprioception. They are rarely used and can provide valuable additional information within the recognition tasks. This article presents an analysis of several methods of fusion of perceptual and auditory modalitites. It relies on the use of a perspective camera and a microphone on a moving object recognition problem. Experimental data are also provided on a database of audio/visual objects including cases of visual occlusions and audio corruptions.

Speech recognition systems usually need a feature extraction stage aiming at obtaining the best signal representation. State of the art speaker verification systems are based on cepstrals features like MFCC, LFCC or LPCC. In this article, we propose to use a genetic algorithm to provide new features able to complete the LFCC's. We present an adaptation of the common LFCC feature extractor which consists in designing a filter bank, optimized for a high level fusion issue. Experiments are carried out using a state of the art speaker verification system. Results show that the proposed method improves the system performances on the 2006 Nist SRE Database.

Measuring human movement in real time is of primary importance in a number of new applications of interactive systems and human centered robotics. A major difficulty in this field arises from the high joint redundancy of the human kinematics. Conventional approaches address this problem by installing a rather large number of sensors or markers on the subject. On the other hand, well admitted theories in neurosciences claim that joint synchronization in human movements is governed by so-called synergies, that can be viewed as joint patterns corresponding to a given gesture of a given individual. In this paper, we intend to exploit this property in order to reduce the number of sensors to be installed on a human subject when tracking his/her motion. Namely, our suggested method comprizes two phases. In a learning stage, the subject is asked to complete a given gesture a few times, while he/she is equipped with sensors able to measure his/her full posture. An algorithm is thus used in a second phase to reduce the required number of sensors while reconstructing the whole posture of the subject. Experimental evidence is provided for the particular motion of sit-to-stand transfer, in a study that involves healthy subjects.

The aim of the work presented in this paper is to realize the model for the control of a robotic interface for equilibrium assistance during sit-to-stand transfer. Interactive robotic devices designed as human-centered robotics, can give more comfortable and more efficient solutions than traditional technical devices. One supposes the need of a virtual model of the pathology. This model, called observer, aims at being used in the smoothing control part of this assisting device. A useful property of this observer should be a postural prediction ability. This article presents a study of different neural solutions : a Neural Predictive Observer (NPO) and a Reduced Neural Predictive Observer (RNPO). Records used for the learning have been done from healthy people that stand up normally and quickly. Some tests will also be done in patients with cerebellar syndrome disease. The presented experimental results show the good accuracy of these approaches whatever the speed of the movement.

Speech recognition systems usually need a feature extraction stage aiming at obtaining the best signal representation. In this article we propose to use genetic algorithms to design a feature extraction method adapted to the speaker diarization task. We present an adaptation of the common MFCC feature extractor which consists in designing a filter bank, with optimized bandwidths. Experiments are carried out using a state-of-the-art speaker diarization system. The proposed method outperforms the original filter bank based on the Mel scale one. Furthermore, the obtained filter bank reveals the importance of some specific spectral information for speaker recognition

In this paper we propose a new feature extraction algorithm based on nonlinear prediction: the Neural Predictive Coding model which is an extension of the classical LPC one. This model is applied to speaker verification by the Arithmetic-Harmonic Sphericity (AHS) method. Two different initialization methods are proposed for the coding method based on the Neural Predictive Coding (NPC): classical neural networks initialization and linear initialization. The first model obtains smaller rates. For the linear initialization, we obtain significant improvement in comparison to the most used methods (LPCC, MFCC). This study opens a new way towards different feature extraction schemes that offers better accuracy on speaker recognition tasks.

In this paper, we present a new method of language detection. This method is based on language pair discrimination using neural networks as classifier of acoustic features. No acoustic decomposition of the speech signal is needed. We present an application of our method for the detection of 11 languages for a signal duration of 3 seconds (OGI MLTS corpus). The obtained results highlight scores ranging from 75.1% to 82% according to the language processed.

We present an architecture called the modular neural predictive coding architecture (Modular NPC). The Modular NPC is used for discriminative feature extraction (DFE). It provides an architecture based on phonetics knowledge applied to phoneme recognition. The phonemes are extracted from the Darpa-Timit speech database. Comparisons with coding methods (LPC, MFCC, PLP) are presented: they put in obviousness an improvement of the recognition rates.

We present a new speech coding, named neural predictive coding, based on an original neural predictive structure. The system is designed to predict the samples of a speech signal window from previous ones. The goal of this coding is to extract the signal discriminative features relative to the database from which it is extracted. After a precise description of our coding, we discuss about the coding phase of the algorithm. We present experiences which show an improvement on telephone speech recognition compared with other predictive or frequency coding. The signal windows are extracted from the telephone speech signal database NTIMIT.

The authors study the autoregressive and moving average (ARMA) filter for lidar signal processing. After a short presentation of the atmospheric laser Doppler instrument project (ALADIN), they introduce the objective of this paper, which is to extract the Doppler frequency and to retrieve the spectral width of a noised lidar signal. A general presentation of ARMA filters and parametric adaptive algorithms (PAAs) is provided. Then they present results about the choice of the model, the Doppler frequency estimate, and the spectral width estimate. Finally, they study the possible estimate of SNR, which is biased by the first estimates (Doppler frequency and spectral width).

Proposes an unsupervised model of formal neural networks to learn and recognize binary temporal sequences. In this model, time is represented by its effect on processing and not as an additional dimension of inputs: synaptic efficacy of a connection is the integration time of the signal passing through the connection. The only parameters subject to learning are connection integration times. A local and unsupervised learning of temporal sequences is achieved by assuming that any cell of the network can have a 'spontaneous' activity instead of an only 'evoked' activity as in other models of formal neurons. The only parameters subject to learning are connection integration times. An example of such a network is described and the results of the simulation are discussed.>