The goal of this proposal is twofold. On the one hand we wish to push on fundamental scientific advances on data based modeling of dynamical systems which lay the ground for the design of autonomous/intelligent systems. On the other hand the aim is to devise new artificial intelligent systems that can be applied, e.g., to a new generation of robot human helpers and co-bots in domestic environments. We will pursue these goals through a new synergic and integrative approach that will draw, and merge, concepts and methods from control theory, computational learning and optimization. Our new computational solutions will be motivated and ultimately applied to design the cognitive and control systems of a robot (hereby called ``the agent’’) with novel actuator and sensing mechanisms, including but not limited to tactile and computational vision modules. With a broader perspective, the intelligent systems we have in mind include • Industrial and service robots (e.g. for assisted living and/or medical purposes) • Smart human-machine interfaces and smart environments • smart integrated systems and automation of complex system Key requirements for such systems are: • safely interaction with an unknown and changing environment, possibly including humans • adaption of their behavior to variable conditions while performing certain assigned tasks in an efficient and robust manner (w.r.t. to uncertainties and changing conditions) . It is fair to say that data drive modeling tools are essential building blocks of autonomous and artificial intelligent systems, which have recently seen some strikingly applications. For example, mobile phones can now be controlled through speech and cars can break on the basis of what they see. Other examples are the new video game consoles, language processing systems such as Watson, as well as a variety of Google services such as translation and visual search. These stories mark the first exciting successes of artificial intelligence and push the frontiers and the challenges of the field. Modeling has been extensively studied in the control community and, quoting from a recent survey by L. Ljung and co-authors,``it has been observed that obtaining the model is the single most time consuming task in the application of model-based control and that three quarters of the total costs associated with advanced control projects can be attributed to modeling. This hints that modeling risks becoming a serious bottleneck in future engineering systems.‘’ In addition, a common trait in the solutions to all the above problems is some degree of ad hoc human intervention often requiring complex and heavy programming. Besides, the amount of the data needed to an artificial system to solve relatively simple tasks seems often very large, with a corresponding increased computational burden. For example, despite the proliferation of robust robotic platforms, commercial robots still rely on a substantial amount of ad hoc engineering, and fall short at performing even simple tasks such as finding, recognizing, tracking and manipulate objects, or moving across an unknown environment. As a consequence, robots cannot handle many of the everyday tasks we would like them to perform. They have a limited ability to safely interact with and navigate through both simple domestic areas as well as more complicated outdoor environments where their potential use for performing repetitive or dangerous tasks, helping to extend human abilities and increase safe mobility, is enormous. Future society will be older and could count on more intelligent, efficient and easy to use robots. As a consequence a truly intelligent system capable to perform a complex task, requiring the solution not of one but of several problems, in an unknown, complex and time-varying environment, is the next natural, yet ambitious challenge. The prototype problem we have in mind is that of a robot that can be instructed to safely and accurately detect, identify and manipulate objects, in a domestic yet unstructured environment. More precisely this encompasses at least the following two steps: • develop new tools for dynamic modeling in a time-varying setup, possibly dealing with high dimensional data; these tools will allow adaptation to changing conditions. A key ingredient to adaptation is the capability of trading task exploitation and learning. This in turn requires endowing the learning mechanism with an accurate uncertainty description around the estimated model. More specifically, on line learning can be made possible only through interaction with the environment (i.e. by acquiring new data through a variety of sensors such as vision, radar, accelerometers, tactile etc.) and has to be done taking into account and trading off the available resources (computation, energy), the expected improvement in the model (how much ``useful’’ information will the agent gather) as well as the assigned tasks (can a task ``wait’’? does the agent need to improve his model to increase his chance of success?). integrate deterministic and statistical techniques to control an actuator with a recently proposed novel hardware design, integrating the sensor and motor systems. The core of our research will be devoted to develop new and effective computational methods, that can provably solve the above problems. We believe that the key towards provably solving such complex issues is the integration of results and methods from control theory and machine learning, together with a new computational architecture exploiting the latest, as well as new, solutions in optimization theory. In fact, very recently, machine learning ideas have been successfully integrated into system identification, proving their potential. One of the goals of this proposal is to fully develop these preliminary ideas. This has led to establishing a consortium which brings together expertise from control/system identification (Padova), machine learning (Genova) and optimization (Ferrara). These three Italian units are also supported by the associated partners: Linkoping University, University of Washington, Seattle, and MIT, Boston. The prototype robotic testbed will be developed in collaboration with Dr. F. Nori and his group at the Robotics, Brain and Cognitive Sciences Department, Italian Institute of Technology (IIT), Genova. Given the interdisciplinary nature, we have decided to articulate the proposal in 4 years, mainly motivated by the need to preliminary establish a solid background common to the three units. To this purpose, in the spirit of the ``Futuro in Ricerca’’ call, we have decided to include three researcher positions (one per unit) for a duration of 3 years. This would allow to strengthen the three units (which are lead by young faculty members) with qualified personnel able to further establish a sustainable and fruitful collaboration among the three scientific areas.
Bando FIRB 2012 Programma Futuro in Ricerca - Protocollo RBFR12M3AC - Apprendere nel tempo:un nuovo approccio computazionale per l'apprendimento automatico di sistemi dinamici - Learning meets time: a new computational approach for learning in dynamic systems
BONETTINI, Silvia;
2012
Abstract
The goal of this proposal is twofold. On the one hand we wish to push on fundamental scientific advances on data based modeling of dynamical systems which lay the ground for the design of autonomous/intelligent systems. On the other hand the aim is to devise new artificial intelligent systems that can be applied, e.g., to a new generation of robot human helpers and co-bots in domestic environments. We will pursue these goals through a new synergic and integrative approach that will draw, and merge, concepts and methods from control theory, computational learning and optimization. Our new computational solutions will be motivated and ultimately applied to design the cognitive and control systems of a robot (hereby called ``the agent’’) with novel actuator and sensing mechanisms, including but not limited to tactile and computational vision modules. With a broader perspective, the intelligent systems we have in mind include • Industrial and service robots (e.g. for assisted living and/or medical purposes) • Smart human-machine interfaces and smart environments • smart integrated systems and automation of complex system Key requirements for such systems are: • safely interaction with an unknown and changing environment, possibly including humans • adaption of their behavior to variable conditions while performing certain assigned tasks in an efficient and robust manner (w.r.t. to uncertainties and changing conditions) . It is fair to say that data drive modeling tools are essential building blocks of autonomous and artificial intelligent systems, which have recently seen some strikingly applications. For example, mobile phones can now be controlled through speech and cars can break on the basis of what they see. Other examples are the new video game consoles, language processing systems such as Watson, as well as a variety of Google services such as translation and visual search. These stories mark the first exciting successes of artificial intelligence and push the frontiers and the challenges of the field. Modeling has been extensively studied in the control community and, quoting from a recent survey by L. Ljung and co-authors,``it has been observed that obtaining the model is the single most time consuming task in the application of model-based control and that three quarters of the total costs associated with advanced control projects can be attributed to modeling. This hints that modeling risks becoming a serious bottleneck in future engineering systems.‘’ In addition, a common trait in the solutions to all the above problems is some degree of ad hoc human intervention often requiring complex and heavy programming. Besides, the amount of the data needed to an artificial system to solve relatively simple tasks seems often very large, with a corresponding increased computational burden. For example, despite the proliferation of robust robotic platforms, commercial robots still rely on a substantial amount of ad hoc engineering, and fall short at performing even simple tasks such as finding, recognizing, tracking and manipulate objects, or moving across an unknown environment. As a consequence, robots cannot handle many of the everyday tasks we would like them to perform. They have a limited ability to safely interact with and navigate through both simple domestic areas as well as more complicated outdoor environments where their potential use for performing repetitive or dangerous tasks, helping to extend human abilities and increase safe mobility, is enormous. Future society will be older and could count on more intelligent, efficient and easy to use robots. As a consequence a truly intelligent system capable to perform a complex task, requiring the solution not of one but of several problems, in an unknown, complex and time-varying environment, is the next natural, yet ambitious challenge. The prototype problem we have in mind is that of a robot that can be instructed to safely and accurately detect, identify and manipulate objects, in a domestic yet unstructured environment. More precisely this encompasses at least the following two steps: • develop new tools for dynamic modeling in a time-varying setup, possibly dealing with high dimensional data; these tools will allow adaptation to changing conditions. A key ingredient to adaptation is the capability of trading task exploitation and learning. This in turn requires endowing the learning mechanism with an accurate uncertainty description around the estimated model. More specifically, on line learning can be made possible only through interaction with the environment (i.e. by acquiring new data through a variety of sensors such as vision, radar, accelerometers, tactile etc.) and has to be done taking into account and trading off the available resources (computation, energy), the expected improvement in the model (how much ``useful’’ information will the agent gather) as well as the assigned tasks (can a task ``wait’’? does the agent need to improve his model to increase his chance of success?). integrate deterministic and statistical techniques to control an actuator with a recently proposed novel hardware design, integrating the sensor and motor systems. The core of our research will be devoted to develop new and effective computational methods, that can provably solve the above problems. We believe that the key towards provably solving such complex issues is the integration of results and methods from control theory and machine learning, together with a new computational architecture exploiting the latest, as well as new, solutions in optimization theory. In fact, very recently, machine learning ideas have been successfully integrated into system identification, proving their potential. One of the goals of this proposal is to fully develop these preliminary ideas. This has led to establishing a consortium which brings together expertise from control/system identification (Padova), machine learning (Genova) and optimization (Ferrara). These three Italian units are also supported by the associated partners: Linkoping University, University of Washington, Seattle, and MIT, Boston. The prototype robotic testbed will be developed in collaboration with Dr. F. Nori and his group at the Robotics, Brain and Cognitive Sciences Department, Italian Institute of Technology (IIT), Genova. Given the interdisciplinary nature, we have decided to articulate the proposal in 4 years, mainly motivated by the need to preliminary establish a solid background common to the three units. To this purpose, in the spirit of the ``Futuro in Ricerca’’ call, we have decided to include three researcher positions (one per unit) for a duration of 3 years. This would allow to strengthen the three units (which are lead by young faculty members) with qualified personnel able to further establish a sustainable and fruitful collaboration among the three scientific areas.I documenti in SFERA sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
