Understanding infant development through baby humanoid robots

Human newborns are at first virtually helpless, but within a year or two they learn to effectively control their bodies, to reach and to locomote. What brain algorithms are at the basis of this learning? Current knowledge on sensorimotor development is patchy. A common but unverified idea is that learning involves intrinsically motivated exploration of the body and the world, giving rise to coherent internal representations that subsequently allow effective motor control. In this talk, I will question these assumptions.

We longitudinally follow infants in three different scenarios - spontaneous behavior, reaching for tactile stimuli on the body, and reaching for objects presented visually - and use behavioral data from these three contexts to: (i) identify signatures of active exploration and goal-orientedness in spontaneous behavior, (ii) compare reaching to somatosensory and visual targets over development to understand the interplay of target localization and motor control. 

To shed light on the mechanisms of the development of the "sensorimotor self", we instantiate these scenarios in baby humanoid robots. I will also show AI tools that we develop to extract infant 3D pose from videos, reattribute motion data to robots, and replay the motor, proprioceptive, visual, and tactile experience of infants. 

Explainable Robotic Systems in Reinforcement Learning Scenarios

Robotic systems are more present in our society everyday. In human-robot environments, it is crucial that end-users may correctly understand their robotic team-partners, in order to collaboratively complete a task. To increase action understanding, users demand more explainability about the decisions by the robot in particular situations.

Recently, explainable robotic systems have emerged as an alternative focused not only on completing a task satisfactorily but also on justifying, in a human-like manner, the reasons that lead to making a decision. In reinforcement learning scenarios, a great effort has been focused on providing explanations using data-driven approaches, particularly from the visual input modality in deep learning-based systems. This talk focuses rather on the decision-making process of reinforcement learning agents performing a task in robotic scenarios.

Intrinsically Motivated Open-ended Learning in Autonomous Agents

Open-ended learning refers to the capacity of an agent to autonomously and continually acquire knowledge and skills, without relying on an exhaustive external specification of tasks or objectives. This capacity is central to the development of both biological and artificial agents able to operate in dynamic, unstructured, and potentially unknown environments, where the relevant competencies cannot be fully anticipated at design time. Drawing on developmental learning and cognitive science, intrinsic motivations offer a principled mechanism to support such processes: self-generated signals that drive agents to explore novel states, discover and select their own goals, and progressively learn associated skills, independently of any task-specific reward.

This talk examines the role of intrinsic motivations in autonomous open-ended learning. It introduces the notion of open-ended learning and its relevance for developmental robotics, discusses how different forms of intrinsic motivation support autonomous exploration, goal discovery and skill learning, and considers how these mechanisms can be implemented, also in relation to other learning approaches. Particular attention is given to the integration of motivational signals within cognitive architectures, where the management of motivations, the storage of acquired knowledge, and the learning of new behaviours must be coordinated and balanced. The talk aims to provide both a conceptual framing and concrete illustrations of how intrinsic motivations can be inserted into architectures for autonomous learning, and to highlight open challenges for the field.