Keywords
Autonomous robotics and development with humanoid robots; computational embodied neuroscience; non-supervised, supervised, and reinforcement learning; open-ended learning; intrinsic and extrinsic motivations; goal-directed and habitual behavior; attention and active vision; hierarchical sensorimotor architectures; Pavlovian and instrumental learning; brain: amygdala, hippocampus, basal ganglia, cerebellum, cortex.
Informal history (a bit long) of my research
These web pages mainly concern my scientific research activities. Since the history of things and people explains a lot of them, here is the informal story of my training and my research experiences.
As high school I did the Italian Scientific High School (at "Cavour" School, Rome, 1987). Then I studied Economics at the University of Rome Sapienza with the idea of doing something useful for people (but maybe I should have studied physics as I was very interested in such topic).
During the university I started attending some courses at the Department of Philosophy, University of Rome Sapienza, as I have always been interested in understanding knowledge and intelligence, the process that produces it. Following some courses, however, I felt that philosophy treated these problems in a way that was not sufficiently scientific (at least in Rome) and so subsequently I followed courses in the Department of Psychology, University of Rome Sapienza. Here, thanks to the fascinating lessons of Prof. Eliano Pessa, I learned the computational approach to the study of intelligence and in particular I was "struck" by neural networks, the model of artificial intelligence that best represents "learning", intended as a process of generation and acquisition of knowledge, a theme on which I have ever since dedicated my research (starting from the final research thesis in Economics where I simulated with neural networks a group of agents forming an oligopolistic market).
So, after graduating in Economics (1997) I a one-year Specialization School in "Neural Networks and Cognitive Psychology" at the Deparment of Psychology, University of Rome Sapienza. In the meantime I started attending the Institute of Cognitive Sciences and Tegnologies of the Italian National Research Council (ISTC-CNR) where I started collaborating with Domenico Parisi, a visionary researcher who contributed to bring neural networks to Italy.
Later (1998) I started a three-year "PhD in Computer Science" at the University of Essex (Colcester, United Kingdom) where I focused on "planning with neural networks and reinforcement learning" under the guidance of Prof. Jim Doran, one of the researchers who contributed to early developments in Classical Artificial Intelligence in Britain. During the PhD reinforcement learning (learning "by trial and error") became a main instrument of investigation in my research. Before, during and after my doctorate, I continued to study psychology and neuroscience in depth.
After the PhD (2001), I returned to Rome at the ISTC-CNR where, thanks to Stefano Nolfi, one of the most important researchers of "Evolutionary Robotics", I made postdocs in some European projects (Swarm-bots , ECAgents). Here my research involved "embodied systems" of "artificial life" (evolutionary approaches with genetic algorithms) for the study of self-organised social phenomena through collective robotic systems. This research reinforced in me the idea that "true" intelligence can only emerge from the interaction of agents with the environment through a body (sensors and actuators).
Later I tried to combine this "embodied vision" of intelligence with the study of the brain and behavior through computational models in a way more strongly linked to the empirical data coming from psychology and neurosciences, in particular working in other European projects within which I started working independently (MindRaces, ICEA). In 2006 I created a research group at the ISTC (where I became Researcher in 2009) called "LOCEN -Laboratory of Computational Embodied Neuroscience" dedicated to studying behavior and brain through computational models that make a synthesis between "embodied vision" of intelligence and empirical constraints coming from neuroscience and psychology data. With LOCEN we progressively specialized in studying the quintessential forms of learning, those related to "curiosity" (intrinsic motivations: surprise, novelty, self-generation of objectives, acquisition of competence) as those you see in playing children and scientists. These approaches and themes have been at the heart of the research of two European projects that I conceived and coordinated for the ISTC: first the European project IM-CLeVeR (Intrinsically Motivated Cumulative Learning Versatile Robots; 2009-2013) and then the European project GOAL- Robots (Goal-based Open-ended Autonomous Learning Robots; 2016-2020). In these projects the themes of open-ended learning have been addressed both from the neuroscientific/psychological point of view (especially in IM-CLeVeR) and the autonomous robotics point of view (specially in GOAL-Robots).
These approaches have led us to deepen both the fundamental mechanisms behind natural learning (neuroscience and psychology: Pavlovian and instrumental learning, goal-directed and habitual behavior, sensorimotor coordination, attention; brain: hierarchical sensorimotor architectures, motivations, amygdala, hippocampus, basal ganglia, cerebellum, cortex) and the mechanisms of learning in artificial intelligence and robotics (developmental and autonomous robotics: vision-manipulation coordination, open-ended learning, intrinsic and extrinsic motivations, active vision; machine learning : unsupervised learning, supervised learning, reinforcement learning, classical and "deep" neural networks).
The main objectives of my research can therefore be grouped into two strands
The first research objective aims at understanding the fundamental mechanisms of human intelligence and behavior, with a particular focus on learning (hence also the strong interest in developmental psychology). This is done through the understanding of the architecture and functioning of the brain system, in particular of how the set of neurons that form it, organized into structures forming complex networks, are able to produce the behavior seen as a circular brain-body-environment interaction ("embodied" view of intelligence). The topic that groups the specific research themes is the cumulative learning of several sensorimotor actions guided by extrinsic motivations (hunger, thirst, survival, etc.) and intrinsic (curiosity: surprise, novelty, competence acquisition, goal self-generation). To this end, in LOCEN we build, on the basis of data from neuroscience, computational models of the brain at the system level: this high level allows us to use the models to understand how the brain generates behavior, an element characterizing all LOCEN studies of brain (the behavior of models is compared with data from psychology). Over the years this has led us to study functions and mechanisms of different areas of the brain (in particular: amygdala, hippocampus, basal ganglia, cerebellum, and different areas of the cortex) and how they interact to produce learning and behavior (classical and instrumental learning), habitual behavior, goal-directed behavior, eye-hand coordination, high-level decision-making and executive processes, and lately consciousness). This also led us to study various mental and neurodegenerative diseases (Parkinson's, Alzheimer's, autism, depression, substance and behavioural dependencies, post-traumatic stress disorder) as trying to understand brain malfunction is not only useful in itself, but it is also one of the main ways to understand the functioning of the healthy brain.
The second objective aims to produce robots able to autonomously learn sensorimotor knowledge and skills in a cumulative way on the basis of intrinsic motivations. In particular, this objective focuses on the development of autonomous robots able to learn to solve many tasks, not just one or a few specific tasks as often happens in robotics. In addition, robots must learn autonomously: from the point of view of applications, this point is very important to limit the use of human work to program the robots, and also to allow the robots to solve tasks in environments that present challenges that cannot be anticipated at the time of their programming (for example, the non-structured environments that are faced by service robots, or the robots that explore new environments such as the space). The strategy to build these robots is "open learning" based on the self-generation of the objectives (tasks) on the basis of intrinsic motivations (novelty, surprise, acquisition of competence): the self-generated goals then allow the robots to acquire motor skills and models of the world. This knowledge can then be used by robots to perform tasks useful for human users. At LOCEN we believe that the development of "true" artificial intelligence must pass through these processes.
