INDEX
- My research intersts: narrative and detailed
- My first set of research objectives and interests: open-ended learning robots
- My second set of research objectives and interests: brain, sensorimotor behaviour, goal-directed cognition and consciousness
- My third set of research objectives and interests: applications of transitional wearables, school in AI, spin-offs AI2Life and HugLab
- Projects on the topics described
1. MY RESEARCH INTERESTS: NARRATIVE AND DETAILED
Since the history of things and people explains much about who they become, here is the informal story of my training, research experiences, and research interests.
High School (1981–1987, Liceo Scientifico Cavour, Rome). I attended an Italian scientific high school, where I cultivated a strong interest in mathematics and the natural sciences.
Economics (1987–1997, BA and MA, University of Rome “La Sapienza”). I initially studied Economics with the idea of doing something useful for society. In retrospect, I sometimes think I should have studied physics, given my deep interest in understanding the general principles governing the world and my passion for mathematics. Nevertheless, economics has proved valuable, as it is lately informing my reflections on how current socio-techno-economic systems might be improved.
University Explorations in Philosophy and Psychology — and Falling in Love with Neural Networks (1992–1997). During my Economics studies, I attended courses in philosophy at La Sapienza because I was drawn to fundamental questions about knowledge and intelligence. Over time, however, I felt that philosophy—at least in that environment—did not address these questions in a sufficiently scientific way, so I shifted toward psychology. There, through the inspiring teaching of Prof. Eliano Pessa, I encountered the computational approach to intelligence and discovered neural networks, which immediately captured me. Neural networks offered the most compelling model of learning—both as knowledge generation and knowledge acquisition—and from that moment onward they became a central focus of my research. My Masters thesis in Economics already reflected this: I built neural-network agents competing in an oligopolistic market, each adapting through genetic algorithms.
In parallel, I also attended courses in neuroscience in the Biology faculty and studied the fundamentals of neuroscience extensively on my own.
First Cognitive Models (1997–1998, Specialization in Neural Networks and Cognitive Psychology, La Sapienza). After graduating in Economics, I completed a one-year Specialization School in Neural Networks and Cognitive Psychology. In this period, I also began collaborating at the Institute of Cognitive Sciences and Technologies (ISTC-CNR), working with Domenico Parisi, a visionary figure who introduced neural network modelling to Italy. With him, I investigated models of human attention and cultural evolution, marking my first steps into computational modelling of cognition.
Planning with Neural Networks (1998–2001, PhD in Computer Science, University of Essex, UK). I then pursued a PhD in Computer Science at the University of Essex under Prof. Jim Doran, one of the pioneers of Classical AI in the UK. My PhD focused on planning with neural networks and reinforcement learning, reinforcing my interest in learning “by trial and error” as a central mechanism in both natural and artificial intelligence. Before, during, and after the PhD, I deepened my knowledge of neuroscience and psychology, aiming to connect computational models with empirical insights.
The Evolutionary Perspective on Intelligence (2001–2004, Postdoctoral Researcher, ISTC-CNR). After the PhD, I returned to ISTC-CNR and worked with Dr. Stefano Nolfi, one of the founders of Evolutionary Robotics. In European projects such as Swarm-bots and ECAgents, I developed embodied systems grounded in the principles of artificial life, complex systems, and evolutionary dynamics (e.g., group selection, genetic algorithms). We studied the emergence of collective behaviours in groups of simple robots, focusing on self-organisation, cooperation, and division of labour. Focusing on Individual Cognition: Attention, Motivation, and Reinforcement Learning (2005–2009, Postdoc, ISTC-CNR). I later sought to integrate the embodied approach with a stronger connection to empirical psychology and neuroscience.
In European projects such as MindRaces and ICEA, I began conducting research more independently, studying attention, motivations and emotions as the mechanisms that drive learning and behaviour. All of this was approached through bio-inspired neural-network models combined with reinforcement learning.
Founder and Principal Investigator of LENAI — Laboratory of Embodied Natural and Artificial Intelligence (2009–present). In 2009, I founded the research group LENAI at ISTC-CNR, becoming a Researcher that year and later Research Director in 2021. LENAI studies the brain and behaviour through system-level computational models that integrate the embodied view of intelligence with constraints from neuroscience and psychology. Over time, we specialised in studying the fundamental forms of learning, particularly those related to curiosity and intrinsic motivations—novelty, surprise, self-generated goals, and competence acquisition—so prominent in infants and children at play.
2. MY FIRST SET OF RESEARCH OBJECTIVES AND INTERESTS: OPEN-ENDED LEARNING ROBOTS
The first set of research objectives aims to develop robots capable of autonomously acquiring sensorimotor knowledge and skills in a cumulative, open-ended fashion through direct interaction with the environment. The overarching goal is to understand the general principles through which autonomous systems acquire knowledge -a central challenge in robotics and a fundamental process in biological organisms.
Our approach is based on open-ended learning, inspired by how infants and children learn (developmental robotics). In this framework, robots self-generate goals and tasks driven by intrinsic motivations, computational analogues of human curiosity. Intrinsic motivations fall into three main families: novelty, surprise, and competence improvement. The goals produced by these mechanisms guide robots in acquiring motor skills and building world models that support behaviour and planning.
Implementing open-ended learning requires the design of sophisticated robot control architectures composed of multiple interacting components, including intrinsic motivation systems, perceptual abstraction modules, goal repertoires, experience memories, hierarchical action structures, world models, and planning processes.
Algorithmically, intrinsic motivations rely on specialised custom algorithms; hierarchical action acquisition is typically driven by reinforcement learning; self-supervised learning is used for acquiring predictive world models; and unsupervised learning (e.g., autoencoders) is used to build perceptual abstractions.
These architectures are tested in both simulated and real robots, with a particular focus on overcoming the challenging sim-to-real transfer problem. A significant difficulty in this field is that open-ended learning experiments often require substantial time for robots to learn autonomously. This makes it necessary to use fast 3D physics simulations and high-performance computing resources.
The knowledge autonomously acquired by these robots can then be reused to accomplish tasks that are meaningful for humans. This is especially crucial when robots must operate in unstructured environments whose challenges cannot be fully anticipated at design time—such as human-inhabited spaces including homes, offices, workplaces, and hospitals.
We are currently investigating several key problems related to open-ended learning. A first challenge concerns the very definition of open-ended learning: how should it be formally characterised so that it can guide the development of concrete robotic solutions? Closely related is the question of how to measure open-ended learning in an objective and reproducible way—an essential requirement for comparing models and advancing the field. We are also exploring how to describe the diverse solutions that open-ended learning systems generate, such as their patterns of autonomous exploration, their trajectories of self-generated goals, their emergent learning curricula, and the behaviours that arise from them. Finally, we aim to clarify the application contexts in which open-ended learning is most beneficial —for example, environments that pose unpredictable challenges or scenarios where users cannot specify all relevant tasks in advance.
Another set of problems concerns the relationships between differnt intrinsic motivations. How can various intrinsic motivation signals be used to drive the learning processes within different robot architectures? How can they be coordinated so that the system’s components develop harmoniously and in a balanced way? And how do intrinsic motivations relate to other promising theoretical concepts—such as free energy minimisation, mutual information, or empowerment —that also aim to formalise curiosity-driven behaviour?
A further challenge involves abstraction. How can we design state-abstraction mechanisms that allow robots to discover latent representations supporting downstream open-ended learning processes, such as world models and action policies? What inference biases or architectural constraints could facilitate the acquisition of such abstractions? Is abstraction itself a core driver of open-ended learning, for example by shaping the emergence of learning curricula?
An additional problem concerns the organisation of acquired knowledge, both for reactive behaviour and for planning and problem solving through world models. How can a robot autonomously decompose action into reusable chunks that can be flexibly recombined to produce increasingly complex behaviours? How should these action chunks be stored and composed? How can we prevent catastrophic forgetting and instead accumulate knowledge in a stable and structured way? And how can decomposition and composition be extended to higher cognitive levels —for instance, within world models that support planning, reasoning, and problem solving?
A final challenge concerns the leap toward higher levels of goal-directed cognitive flexibility. How can we enable AI systems and robots to monitor their current behaviour, evaluate it, and shift toward more abstract or more globally relevant goals when needed? Can mechanisms observed in humans —such as metacognition, introspection, and consciousness— inspire or support such abilities? And what are the distinct functional roles of these processes in cognition, and what computational mechanisms could underpin them in artificial systems?
3. MY SECOND SET OF RESEARCH OBJECTIVES AND INTERSTS: BRAIN, SENSORIMOTOR-BEHAVIOUR, GOAL-DIRECTED COGNITION AND CONSCIOUSNESS
The second set of research objectives aims to understand the fundamental mechanisms of human intelligence and behaviour, with a particular emphasis on learning. This research thread relies on data from neuroscience, which provide insights into the organisation and functioning of the brain, and on data from cognitive and developmental psychology, which inform us about behaviour and cognition across the lifespan.
The topics investigated span a broad range of sensorimotor and cognitive phenomena. These include visual attention, studied through actively controlled small visual sensors; eye–hand coordination during object manipulation; and higher-level cognitive functions such as cognitive flexibility, planning, and problem solving. We are also interested in cognitive monitoring processes related to metacognition, introspection, and consciousness.
This research is carried out through theory-driven computational models, in contrast to data-driven black-box machine-learning systems. Theory-driven models aim not only to reproduce behavioural or cognitive functions, but to do so by constraining their architectural organisation according to neuroscientific knowledge about brain anatomy and physiology. In this way, the models generate behaviour through the same kinds of computational meachanisms that biological systems do.
Within LENAI, our models are typically system-level computational embodied neuroscience models. ‘System-level’ means that the models reproduce the macro-architecture of brain systems underlying the behaviours of interest. They include large-scale structures —such as the major cortical areas and distinct basal ganglia districts—and, when possible, meso-level features such as internal organisation patterns. Only rarely do we model the micro-level neural physiology (e.g., local circuits and STDP), which is far more challenging and often unnecessary for explaining behaviour.
These models are often embodied, meaning they reproduce behaviour and cognition within the circular loop that joins brain, body, and environment. Embodiment emphasises the continuous interaction through which action shapes perception, generating both challenges and opportunities for adaptation. It also highlights the role of bodily needs —energy, safety, social bonds— which strongly influence emotions and motivations.
Because of these properties, our models can yield explanations and predictions not only about behaviour but also about brain organisation and function. Their system-level and embodied nature also fosters synergy with artificial intelligence and machine-learning research, where the emphasis is on functional behaviour rather than detailed neurophysiology. At the same time, this approach poses substantial methodological challenges: ecological behaviour must be studied through interactive experiments, whereas much of data from cognitive psychology relies on simplified stimulus–response paradigms with limited environmental coupling.
The behavioural phenomena we study centre on learning. These include classical and instrumental conditioning, and the underlying habitual and goal-directed systems, as characterised by behavioural psychology. We study sensorimotor learning, in particular eye–hand coordination, described in developmental psychology: in particular, infant sensorimotor behaviour driven by both extrinsic motivations (e.g., hunger, thirst, safety) and intrinsic motivations (curiosity, novelty, surprise, competence acquisition, goal self-generation). We investigate higher-level cognition studied by cognitive psychology, includign planning and priblem solving; and how these ground human highly distinctive capabilities for metacognition, free-will, introspection and consciousness.
We study these behavioural and cognitive processes through theory-driven computational models of the brain systems that support them. A major focus lies on the basal ganglia–cortical loops, including ventral circuits underlying motivational processes and goal selection; medial circuits involved in decision-making and goal setting; and dorsal circuits responsible for sensorimotor control. We view the basal ganglia as specialised in learning to select and gate the contents flowing through cortical pathways. For example, the dorsal pathway supporting object manipulation extends from early visual areas (e.g. V1, V2), through associative parietal regions, to premotor and motor cortex; while the ventral pathway supporting decision-making extends from the visual system to parietal object-recognition areas, and then into prefrontal regions responsible for high-level cognition, goal setting, working memory, planning, and problem solving.
Sensorimotor behaviour is also shaped by the cerebellum, which supports predictive control; while hippocampus supports episodic memory, action sequences, and planning. Goal selection and habit acquisition are modulated by the dopaminergic system, modulated by structures such as the amygdala and nucleus accumbens (crucial for Pavlovian learning, Pavlovian-to-instrumental transfer, and goal-setting), the hypothalamus (generating primary needs, including social needs), and the periaqueductal gray (hosting innate behavioural patterns).
The overarching objective of developing these models is to understand the general principles through which the brain acquires knowledge, processes information, and generates behaviour. I consider this one of the most fascinating topics one can study, as it touches the essence of knowledge, emotion, motivation, intelligence, and ultimately what makes us human.
A further objective is to study how these systems malfunction in diseases and impairments. Understanding impairments allows us to better understand normal functioning, and it offers insights into neurological and psychiatric disorders. We investigate conditions such as Parkinson’s disease, behavioural and substance addictions, and obsessive–compulsive behaviors —conditions linked to dysfunctions in dopaminergic and habit-learning systems. We also study Alzheimer’s disease, which compromises high-level cognitive functions, and developmental disorders, especially autism, which involves stereotyped actions and social impairments.
4. MY THIRD SET OF RESEARCH OBJECTIVES AND INTERESTS: APPLICATIONS OF TRANSITIONAL WEARABLES, SCHOOL IN AI, SPIN-OFFs AI2Life AND HugLab
I am also deeply interested in the possible applications of the knowledge gained through the investigations described above.
A first application concerns the development of wearable devices, plush-like interactive robots that can be used to monitor, study, and support therapeutic interventions for children with developmental disorders, particularly autism. These wearables provide customisable stimuli —such as lights and sounds— that therapists can tailor to each child in order to elicit intrinsic motivations and via these foster social interaction, an area where children with autism often experience difficulties. Over the years, we have physically developed several devices of this kind, including Plus-Me and Octopus. In 2025, together with LENAI, we founded a CNR spin-off start-up, HugLab, dedicated to producing wearables on a larger scale and offering support to clients such as therapy centres. These devices can also be used as toys for social and emotional education in elementary schools.
A second application concerns teaching and training in the fields outlined above. To this end, in 2018 I founded, and served as President of, the AS-AI Advanced School in Artificial Intelligence. I created the School when it became clear that an “AI tsunami” was approaching and that society urgently needed accessible, high-quality training. AS-AI is a postgraduate program of roughly 300 hours spread over one year, based on synchronous online lectures, hands-on sessions, and a final applied project. The School trained around 25 postgraduate students per year, coming from diverse academic backgrounds—ranging from STEM disciplines to the humanities.
In 2025, however, the School was suspended because its target audience —postgraduates and professionals needing upskilling— struggled to cover tuition fees, despite the fact that the value of the School for students’ careers and lives far exceeded its cost. For this reason, we are now seeking public funding to support and relaunch AS-AI. Public investment in AI education has never been more urgent: it is essential to ensure inclusion, protect jobs, and enable society to adapt to rapid technological change.
In 2021, I also initiated the creation of, and became a partner in, another CNR spin-off start-up, AI2Life, established to manage AS-AI. The start-up is also developing AI applications, such as chatbots for universities, and is preparing to address the imminent and widespread integration of robots into daily human life, a development expected to grow in parallel, and in synergy, with the recent advances in artificial intelligence.
5. PROJECTS ON THE TOPICS ABOVE
The topics described above have led me to propose, coordinate, or serve as Principal Investigator for several European and national projects. Below I report the most significant ones for my research (see the Projects tab of this webpage for further details):
- 2004–2007 – Team Leader: MindRACES: from Reactive to Anticipatory Cognitive Embodied Systems
This project aimed to study the role of anticipation and planning in cognition and in robots. 2006–2009 – Team Leader: ICEA – Integrating Cognition, Emotion and Autonomy
This project employed computational models and embodied agents inspired by the brain and behaviour of rats to investigate the mechanisms and roles of emotions in animals and humans.
2009–2013 – Coordinator: IM-CLeVeR – Intrinsically Motivated Cumulative Learning Versatile Robots
The project focused on developing open-ended robot learning systems by studying intrinsic motivations and sensorimotor learning in humans and animals.
2016–2021 – Coordinator: GOAL-Robots – Goal-based Open-ended Autonomous Learning Robots
This project advanced the development of open-ended learning robots using techniques from machine learning to introduce in architectures goals the pivot allowing intrinsic motivations to support skilsl and knowledge acquisition.
2020–2022 – Coordinator: IM-TWIN – From Intrinsic Motivations to Transitional Wearable Intelligent Companions for Autism Spectrum Disorder
The project developed transitional wearable systems to study autism and support therapeutic interventions for children with Autism Spectrum Disorder.
2021–2023 – Coordinator: GROW – General-purpose Robot for Object Retrieval in Warehouses
This project explored applications of open-ended learning robots in practical domains such as object retrieval in the retail sector.
2022–2026 – Team Leader: PILLAR – Purposeful Intrinsically-motivated Lifelong Learning Autonomous Robots
This project investigates how to use purpose —the broad, high-level goals of human users— to guide and bias open-ended learning processes in robots and make them increasingly relevant for real-world applications.
2023–2026 – Principal Investigator: EBRAINS-Italy – European Brain ReseArch INfrastructureS–Italy
This project aims to create a national infrastructure of services to support research in neuroscience, computational neuroscience, and neurorobotics.
