We propose a system able to represent others' actions as well as to internally simulate their likely continuation from a partial observation. The approach presented here is the first step towards a more ambitious goal of endowing an artificial agent with the ability to recognise and predict others' intentions. Our approach is based on the associative self-organising map, a variant of the self-organising map capable of learning to associate its activity with different inputs over time, where inputs are processed observations of others' actions. We have evaluated our system in two different experimental scenarios obtaining promising results: the system demonstrated an ability to learn discriminable representations of actions, to recognise novel input, and to simulate the likely continuation of partially seen actions.

Understanding and efficiently representing skills is one of the most important problems in a general Programming by Demonstration (PbD) paradigm. We present Growing Hierarchical Dynamic Bayesian Networks (GHDBN), an adaptive variant of the general DBN model able to learn and to represent complex skills. The structure of the model, in terms of number of states and possible transitions between them, is not needed to be known a priori. Learning in the model is performed incrementally and in an unsupervised manner.

Imitation learning has been recognized as a promising technique to teach robots advanced skills. It is based on the idea that robots could learn new behaviors by observing and imitating the behaviors of other skilled actors. We propose an adaptive probabilistic graphical model which copes with three core issues of any imitative behavior: observation, representation and reproduction of skills. Our model, Growing Hierarchical Dynamic Bayesian Network (GHDBN), is hierarchical (i.e. able to characterize structured behaviors at different levels of abstraction), and growing (i.e. skills are learned or updated incrementally - and at each level of abstraction - every time a new observation sequence is available). A GHDBN, once trained, is able to recognize skills being observed and to reproduce them by exploiting the generative power of the model. The system has been successfully tested in simulation, and initial tests have been conducted on a NAO humanoid robot platform.

Language is among the most fascinating and complex cognitive activities that develops rapidly since the early months of infants' life. The aim of the present work is to provide a humanoid robot with cognitive, perceptual and motor skills fundamental for the acquisition of a rudimentary form of language. We present a novel probabilistic model, inspired by the findings in cognitive sciences, able to associate spoken words with their perceptually grounded meanings. The main focus is set on acquiring the meaning of various perceptual categories (e.g. red, blue, circle, above, etc.), rather than specific world entities (e.g. an apple, a toy, etc.). Our probabilistic model is based on a variant of multi-instance learning technique, and it enables a robotic platform to learn grounded meanings of adjective/noun terms. The systems could be used to understand and generate appropriate natural language descriptions of real objects in a scene, and it has been successfully tested on the NAO humanoid robotic platform.

Old photos may be affected by several types of defects. Manual restorers use their own taxonomy to classify damages by which a photo is affected, in order to apply the proper restoration techniques for a specific defect. Once a photo is digitally acquired, defects become part of the image, and their aspect change. This paper wants to be a first attempt to correlate real defects of printed photos, and digital defects of their digitized versions. A dual taxonomy is proposed, for real and digital defects, and used to classify an image dataset, for a posteriori comparative study. Furthermore, a set of digital features is analyzed for digitized images, to identify which of them could be useful for an automatic inspection method.

In this paper we propose a trainable system that learns grounded language models from examples with a minimum of user intervention and without feedback. We have focused on the acquisition of grounded meanings of spatial and adjective/noun terms. The system has been used to understand and subsequently to generate appropriate natural language descriptions of real objects and to engage in verbal interactions with a human partner. We have also addressed the problem of resolving eventual ambiguities arising during verbal interaction through an information theoretic approach.

Digital image restoration aims to recover damaged zones of a digital image, using surrounding information. In this paper we propose a novel approach, based on bit-plane slicing decomposition, with the purpose to make information analysis and reconstruction process easy, fast and effective. Tests have been made on digitized damaged old photos to restore several classes of typical defects in old photographic prints.

Grasping is an essential prerequisite for an agent, either human or robotic, to manipulate various kinds of objects present in the world. It is a fact that we would like robots to have the same skills as we do. However, despite the construction of human-hand-like robotic effectors, much work is still to be done in order to give robots the capability to grasp and manipulate objects. The goal of this work is to automatically perform grasp synthesis of unknown planar objects. In other words, we must compute points on the object's boundary to be reached by the robotic fingers such that the resulting grasp, among infinite possibilities, optimizes some given criteria. The space of possible configurations is analyzed using genetic algorithms. However, genetic algorithms are not suitable to applications where time is a critical issue. In order to achieve real-time characteristics of the algorithm, neural networks are used: a huge training-set is collected off-line using genetic algorithms, and a feedforward network is trained on these values. We will demonstrate the usefulness of this approach in the process of grasp synthesis, and show the results achieved on an anthropomorphic arm/hand robot.

This paper presents the cognitive architecture Con-SCIS (conceptual space based cognitive imitation system), which tightly links low-level data processing with knowledge representation in the context of imitation learning. We use the word imitate to refer to the paradigm of program-level imitation: we are interested in the final effects of actions on objects, and not on the particular kinematic or dynamic properties of the motion. The same architecture is used both to analyze and represent the task to be imitated, and to perform the imitation by generalizing in novel and different circumstances. The implemented experimental scenario is a simplified two-dimensional world populated with various objects in which observation/imitation takes place. During the observation phase, the user shows her/his hand while performing arbitrary tasks of manipulating objects in front of a single calibrated camera. The task is then segmented into meaningful units and its properties (objects' color and shape, their absolute position and orientation, relations between objects) are represented into high-level symbolic terms. In the imitation phase, the symbolic information is employed to drive the robot's actions. To validate our approach, we report some results concerned with the problem of teaching a humanoid hand/arm robotic system tasks of assembling different workspace objects

This paper presents ConSCIS, conceptual space based cognitive imitation system, which tightly links low-level data processing with knowledge representation in the context of robot imitation. Our focus is on the program-level imitation: we are interested in the final effects of actions on objects, and not on the particular kinematic or dynamic properties of the motion. The same architecture is used both to analyze and represent the task to be imitated, and to perform the imitation by generalizing in novel and different circumstances. The implemented experimental scenario is a two dimensional world populated with various objects in which observation/imitation takes place. To validate our approach, we report some results concerned with the problem of teaching a humanoid hand/arm robotic system tasks of assembling different workspace objects

This paper deals with the design and implementation of a visual control of a robotic system composed of a dexterous hand and video camera. The aim of the proposed system is to reproduce the movements of a human hand in order to learn complex manipulation tasks or to interact with the user. A novel algorithm for robust and fast fingertips localization and tracking is presented. A suitable kinematic hand model is adopted to achieve a fast and acceptable solution to an inverse kinematics problem. The system is part of a cognitive architecture for posture learning that integrates the perceptions by a high-level representation of the scene and of the observed actions. The anthropomorphic robotic hand imitates the gestures acquired by the vision system in order to learn meaningful movements, to build its knowledge by different conceptual spaces, and to perform complex interactions with the human operator.

The paper deals with the design and implementation of a visual control of a robotic system composed of a dexterous hand and stereo cameras. The aim of the proposed system is to reproduce the movements of a human hand in order to learn complex manipulation tasks. A novelty algorithm for a robust and fast fingertips localization and tracking is presented. Moreover, a simulator is integrated in the system to give useful feedbacks to the users during operations, and provide robust testing framework for real experiments.