Artificial Intelligence, Robotics, and Computer Vision

Research focus

The Peer review has evaluated this group as Excellent

The main tracks within Artificial Intelligence where the group focuses his efforts are: autonomous agents, machine learning, modeling complex systems by knowledge-based and computational intelligence techniques. People in the group have defined a framework of agent-to-agent interaction as a semantic model of communicative acts that relies on the concept of social commitment. This perspective was shown to achieve results missing in the mainstream view of agent communication, based on mental states. We also deal with the semantic model of a repository of learning objects, which is meant to be the first step towards the use of ontologies for planning activities in e-learning. The adaptive planning and scheduling systems for multi-agent systems, developed in the group, have been successfully adopted in ambient intelligence and space applications. A methodology that uses cooperative negotiation techniques to model complex phoenomena as interacting agents, is applied to control biological systems. In the Genetic-Based Machine Learning (GBML) area, worldwide reference persons for XCS (eXtended Classifier System) are operating in the group. They obtained results concerning the generalization capabilities of these models, the formulation of formal frameworks to model GBML methods from a Reinforcement Learning (RL) perspective, and to define formal properties of these models. The main achievements in the RL area are: an algorithm for the approximation of the utility function in large probabilistic environments, a novel approach for intrinsically motivated learning, an algorithm that deals with the negative effects of the discretization of a continuous state space, a new approach for knowledge transfer among problems. Most of these have been integrated in a versatile learning framework that is used in many different applications. It has successfully participated to international competitions and a benchmark modeled with it has been accepted as part of the world benchmarking system for RL. The group has also used other learning methods in different applications, becoming a reference source w.r.t. the local research community: bayesian approaches to model adaptation and learning, neural models for biological signals, adaptive color models, augmented and alternative language models for user support systems, adaptive models for traffic prediction, optimization and modeling. Some of us have developed methods and tools to support discovery in life sciences by computational models. We successfully approached the area of toxicity prediction, the first stage of risk assesment for health and environment. We have applied many modeling tools to represent knowledge for activity prediction of chemical compounds. We have also developed methods, and promoted the birth and growth of a community able to propose successful “in silico” methods to authorities, as opposed to the more expensive and life-destroying “in vitro” methods for risk assessment. We promoted the development and diffusion of fuzzy modeling for control and decision making, implementing many real applications for robots, expert systems, user modeling, data classification. In the Robotics area, the group has developed autonomous robots since the early Eighties, exploiting: mechanical/electronical design and implementation, programming, sensor data interpretation, world modeling, control architecture, planning, behavior and strategy management. Among the robots developed in the last years, we mention biomimetic robots (humanoid arm and legs, quadrupeds and hexapods) and a team of wheeled Robocup soccer robots, used also for 63 other service robot applications. We implemented a framework for component reuse based on a conceptual model able to offer conceptual information to modules through a publish/subscribe middleware. This is one of the keypoints requested by companies to implement autonomous robots for the market. Metrical world modeling for autonomous robotics has been addressed both in 2D and 3D to solve classical issues such as map building, localization, and exploration. We developed methods to integrate perceptions while exploring unknown environments by their geometrical features, without using any additional information. This has been done with laser scanner and 2D segment-based maps with no pose information, and also, in 3D, using a segment-based model exploiting a 6 degree of fredom robot pose and a proper statistical modeling of visual measurements. We have also established a general approach based on multi-objective optimization to define efficient exploration strategies. In the image analysis area, the group focuses on: feature detection under image blur, edge localization with sub-pixel accuracy, motion segmentation and target tracking, gesture recognition for video-surveillance, visual environment monitoring, color image analysis and adaptive color modeling. Research in computer vision addresses: object recognition and localization, reconstruction of 3D scenes, reconstruction of deformable objects and paper-like surfaces, analysis of motion blur to retrieve 3D motion, visual analysis of sports events, robot visual odometry, self-localization and mapping, catadioptric camera calibration and 3D reconstruction from omni-directional images.

Dipartimento di afferenza

Dipartimento di Elettronica e Informazione (DEI)

Docenti afferenti

Andrea Bonarini (full professor)
Marco Colombetti (full professor)
Vincenzo Caglioti (associate professor)
Giuseppina Gini (associate professor)
Pier Luca Lanzi (associate professor)
Francesco Amigoni (assistant professor)
Nicola Gatti (assistant professor)
Matteo Matteucci (assistant professor)
Piera Sassaroli (assistant professor)