Gianluca Baldassarre

Olivier Sigaud, Gianluca Baldassarre, Cedric Colas et al.

A lot of recent machine learning research papers have ``open-ended learning'' in their title. But very few of them attempt to define what they mean when using the term. Even worse, when looking more closely there seems to be no consensus on what distinguishes open-ended learning from related concepts such as continual learning, lifelong learning or autotelic learning. In this paper, we contribute to fixing this situation. After illustrating the genealogy of the concept and more recent perspectives about what it truly means, we outline that open-ended learning is generally conceived as a composite notion encompassing a set of diverse properties. In contrast with previous approaches, we propose to isolate a key elementary property of open-ended processes, which is to produce elements from time to time (e.g., observations, options, reward functions, and goals), over an infinite horizon, that are considered novel from an observer's perspective. From there, we build the notion of open-ended learning problems and focus in particular on the subset of open-ended goal-conditioned reinforcement learning problems in which agents can learn a growing repertoire of goal-driven skills. Finally, we highlight the work that remains to be performed to fill the gap between our elementary definition and the more involved notions of open-ended learning that developmental AI researchers may have in mind.

Alejandro Romero, Gianluca Baldassarre, Richard J. Duro et al.

Autonomous open-ended learning is a relevant approach in machine learning and robotics, allowing the design of artificial agents able to acquire goals and motor skills without the necessity of user assigned tasks. A crucial issue for this approach is to develop strategies to ensure that agents can maximise their competence on as many tasks as possible in the shortest possible time. Intrinsic motivations have proven to generate a task-agnostic signal to properly allocate the training time amongst goals. While the majority of works in the field of intrinsically motivated open-ended learning focus on scenarios where goals are independent from each other, only few of them studied the autonomous acquisition of interdependent tasks, and even fewer tackled scenarios where goals involve non-stationary interdependencies. Building on previous works, we tackle these crucial issues at the level of decision making (i.e., building strategies to properly select between goals), and we propose a hierarchical architecture that treating sub-tasks selection as a Markov Decision Process is able to properly learn interdependent skills on the basis of intrinsically generated motivations. In particular, we first deepen the analysis of a previous system, showing the importance of incorporating information about the relationships between tasks at a higher level of the architecture (that of goal selection). Then we introduce H-GRAIL, a new system that extends the previous one by adding a new learning layer to store the autonomously acquired sequences of tasks to be able to modify them in case the interdependencies are non-stationary. All systems are tested in a real robotic scenario, with a Baxter robot performing multiple interdependent reaching tasks.

Gianluca Baldassarre, Richard J. Duro, Emilio Cartoni et al.

The unprecedented advancement of artificial intelligence enables the development of increasingly autonomous robots. These robots hold significant potential, particularly in moving beyond engineered factory settings to operate in the unstructured environments inhabited by humans. However, this possibility also generates a relevant autonomy-alignment problem to ensure that robots' autonomous learning processes still focus on acquiring knowledge relevant to accomplish human practical purposes, while their behaviour still aligns with their broader purposes. The literature has only begun to address this problem, and a conceptual, terminological, and formal framework is still lacking. Here we address one of the most challenging instances of the problem: autonomous open-ended learning (OEL) robots, capable of cumulatively acquiring new skills and knowledge through direct interaction with the environment, guided by self-generated goals and intrinsic motivations. In particular, we propose a computational framework, first introduced qualitatively and then formalised, to support the design of OEL robot architectures that balance autonomy and control. The framework pivots on the novel concept of purpose. A human purpose specifies what humans (e.g., designers or users) want the robot to learn, do or not do, within a certain boundary of autonomy and independently of the domains in which it operates.The framework decomposes the autonomy-alignment problem into more tractable sub-problems: the alignment of `robot purposes' with human purposes, either by hardwiring or through learning; the arbitration between multiple purposes; the grounding of purposes into specific domain-dependent robot goals; and the competence acquisition needed to accomplish these goals. The framework and its potential utility are further elucidated through the discussion of hypothetical example scenarios framed within it.

Michele Farisco, Gianluca Baldassarre, Emilio Cartoni et al.

Despite its successes, to date Artificial Intelligence (AI) is still characterized by a number of shortcomings with regards to different application domains and goals. These limitations are arguably both conceptual (e.g., related to underlying theoretical models, such as symbolic vs. connectionist), and operational (e.g., related to robustness and ability to generalize). Biologically inspired AI, and more specifically brain-inspired AI, promises to provide further biological aspects beyond those that are already traditionally included in AI, making it possible to assess and possibly overcome some of its present shortcomings. This article examines some conceptual, technical, and ethical issues raised by the development and use of brain-inspired AI. Against this background, the paper asks whether there is anything ethically unique about brain-inspired AI. The aim of the paper is to introduce a method that has a heuristic nature and that can be applied to identify and address the ethical issues arising from brain-inspired AI. The conclusion resulting from the application of this method is that, compared to traditional AI, brain-inspired AI raises new foundational ethical issues and some new practical ethical issues, and exacerbates some of the issues raised by traditional AI.

Giovanni Granato, Emilio Cartoni, Federico Da Rold et al.

A common view on the brain learning processes proposes that the three classic learning paradigms -- unsupervised, reinforcement, and supervised -- take place in respectively the cortex, the basal-ganglia, and the cerebellum. However, dopamine outbursts, usually assumed to encode reward, are not limited to the basal ganglia but also reach prefrontal, motor, and higher sensory cortices. We propose that in the cortex the same reward-based trial-and-error processes might support not only the acquisition of motor representations but also of sensory representations. In particular, reward signals might guide trial-and-error processes that mix with associative learning processes to support the acquisition of representations better serving downstream action selection. We tested the soundness of this hypothesis with a computational model that integrates unsupervised learning (Contrastive Divergence) and reinforcement learning (REINFORCE). The model was tested with a task requiring different responses to different visual images grouped in categories involving either colour, shape, or size. Results show that a balanced mix of unsupervised and reinforcement learning processes leads to the best performance. Indeed, excessive unsupervised learning tends to under-represent task-relevant features while excessive reinforcement learning tends to initially learn slowly and then to incur in local minima. These results stimulate future empirical studies on category learning directed to investigate similar effects in the extrastriate visual cortices. Moreover, they prompt further computational investigations directed to study the possible advantages of integrating unsupervised and reinforcement learning processes.

Charles Wilmot, Gianluca Baldassarre, Jochen Triesch

A key competence for open-ended learning is the formation of increasingly abstract representations useful for driving complex behavior. Abstract representations ignore specific details and facilitate generalization. Here we consider the learning of abstract representations in a multi-modal setting with two or more input modalities. We treat the problem as a lossy compression problem and show that generic lossy compression of multimodal sensory input naturally extracts abstract representations that tend to strip away modalitiy specific details and preferentially retain information that is shared across the different modalities. Furthermore, we propose an architecture to learn abstract representations by identifying and retaining only the information that is shared across multiple modalities while discarding any modality specific information.

Emilio Cartoni, Davide Montella, Jochen Triesch et al.

Open-ended learning is a core research field of developmental robotics and AI aiming to build learning machines and robots that can autonomously acquire knowledge and skills incrementally as infants and children. The first contribution of this work is to study the challenges posed by the previously proposed benchmark `REAL competition' aiming to foster the development of truly open-ended learning robot architectures. The competition involves a simulated camera-arm robot that: (a) in a first `intrinsic phase' acquires sensorimotor competence by autonomously interacting with objects; (b) in a second `extrinsic phase' is tested with tasks unknown in the intrinsic phase to measure the quality of knowledge previously acquired. This benchmark requires the solution of multiple challenges usually tackled in isolation, in particular exploration, sparse-rewards, object learning, generalisation, task/goal self-generation, and autonomous skill learning. As a second contribution, we present a set of `REAL-X' robot architectures that are able to solve different versions of the benchmark, where we progressively release initial simplifications. The architectures are based on a planning approach that dynamically increases abstraction, and intrinsic motivations to foster exploration. REAL-X achieves a good performance level in very demanding conditions. We argue that the REAL benchmark represents a valuable tool for studying open-ended learning in its hardest form.

Vieri Giuliano Santucci, Davide Montella, Bruno Castro da Silva et al.

When facing the problem of autonomously learning multiple tasks with reinforcement learning systems, researchers typically focus on solutions where just one parametrised policy per task is sufficient to solve them. However, in complex environments presenting different contexts, the same task might need a set of different skills to be solved. These situations pose two challenges: (a) to recognise the different contexts that need different policies; (b) quickly learn the policies to accomplish the same tasks in the new discovered contexts. These two challenges are even harder if faced within an open-ended learning framework where an agent has to autonomously discover the goals that it might accomplish in a given environment, and also to learn the motor skills to accomplish them. We propose a novel open-ended learning robot architecture, C-GRAIL, that solves the two challenges in an integrated fashion. In particular, the architecture is able to detect new relevant contests, and ignore irrelevant ones, on the basis of the decrease of the expected performance for a given goal. Moreover, the architecture can quickly learn the policies for the new contexts by exploiting transfer learning importing knowledge from already acquired policies. The architecture is tested in a simulated robotic environment involving a robot that autonomously learns to reach relevant target objects in the presence of multiple obstacles generating several different obstacles. The proposed architecture outperforms other models not using the proposed autonomous context-discovery and transfer-learning mechanisms.

Manuel Del Verme, Bruno Castro da Silva, Gianluca Baldassarre

Reinforcement learning can greatly benefit from the use of options as a way of encoding recurring behaviours and to foster exploration. An important open problem is how can an agent autonomously learn useful options when solving particular distributions of related tasks. We investigate some of the conditions that influence optimality of options, in settings where agents have a limited time budget for learning each task and the task distribution might involve problems with different levels of similarity. We directly search for optimal option sets and show that the discovered options significantly differ depending on factors such as the available learning time budget and that the found options outperform popular option-generation heuristics.

Giovanni Granato, Gianluca Baldassarre

Goal-directed manipulation of representations is a key element of human flexible behaviour, while consciousness is often related to several aspects of higher-order cognition and human flexibility. Currently these two phenomena are only partially integrated (e.g., see Neurorepresentationalism) and this (a) limits our understanding of neuro-computational processes that lead conscious states to produce flexible goal-directed behaviours, (b) prevents a computational formalisation of conscious goal-directed manipulations of representations occurring in the brain, and (c) inhibits the exploitation of this knowledge for modelling and technological purposes. Addressing these issues, here we extend our `three-component theory of flexible cognition' by proposing the `Goal-Aligning Representations Internal Manipulation' (GARIM) theory of conscious and flexible goal-directed cognition. The central idea of the theory is that conscious states support the active manipulation of goal-relevant internal representations (e.g., of world states, objects, and action sequences) to make them more aligned with the pursued goals. This leads to the generation of the knowledge which is necessary to face novel situations/goals, thus increasing the flexibility of goal-directed behaviours. The GARIM theory integrates key aspects of the main theories of consciousness into the functional neuro-computational framework of goal-directed behaviour. Moreover, it takes into account the subjective sensation of agency that accompanies conscious goal-directed processes (`GARIM agency'). The proposal has also implications for experimental studies on consciousness and clinical aspects of conscious goal-directed behaviour. Finally, the GARIM theory benefit technological fields such as autonomous robotics and machine learning (e.g., the manipulation process may describe the operations performed by systems based on transformers).

Gianluca Baldassarre

There is a growing interest and literature on intrinsic motivations and open-ended learning in both cognitive robotics and machine learning on one side, and in psychology and neuroscience on the other. This paper aims to review some relevant contributions from the two literature threads and to draw links between them. To this purpose, the paper starts by defining intrinsic motivations and by presenting a computationally-driven theoretical taxonomy of their different types. Then it presents relevant contributions from the psychological and neuroscientific literature related to intrinsic motivations, interpreting them based on the grid, and elucidates the mechanisms and functions they play in animals and humans. Endowed with such concepts and their biological underpinnings, the paper next presents a selection of models from cognitive robotics and machine learning that computationally operationalise the concepts of intrinsic motivations and links them to biology concepts. The contribution finally presents some of the open challenges of the field from both the psychological/neuroscientific and computational perspectives.

Angelo Oddi, Riccardo Rasconi, Emilio Cartoni et al.

In symbolic planning systems, the knowledge on the domain is commonly provided by an expert. Recently, an automatic abstraction procedure has been proposed in the literature to create a Planning Domain Definition Language (PDDL) representation, which is the most widely used input format for most off-the-shelf automated planners, starting from `options', a data structure used to represent actions within the hierarchical reinforcement learning framework. We propose an architecture that potentially removes the need for human intervention. In particular, the architecture first acquires options in a fully autonomous fashion on the basis of open-ended learning, then builds a PDDL domain based on symbols and operators that can be used to accomplish user-defined goals through a standard PDDL planner. We start from an implementation of the above mentioned procedure tested on a set of benchmark domains in which a humanoid robot can change the state of some objects through direct interaction with the environment. We then investigate some critical aspects of the information abstraction process that have been observed, and propose an extension that mitigates such criticalities, in particular by analysing the type of classifiers that allow a suitable grounding of symbols.

Vieri Giuliano Santucci, Gianluca Baldassarre, Emilio Cartoni

Autonomous multiple tasks learning is a fundamental capability to develop versatile artificial agents that can act in complex environments. In real-world scenarios, tasks may be interrelated (or "hierarchical") so that a robot has to first learn to achieve some of them to set the preconditions for learning other ones. Even though different strategies have been used in robotics to tackle the acquisition of interrelated tasks, in particular within the developmental robotics framework, autonomous learning in this kind of scenarios is still an open question. Building on previous research in the framework of intrinsically motivated open-ended learning, in this work we describe how this question can be addressed working on the level of task selection, in particular considering the multiple interrelated tasks scenario as an MDP where the system is trying to maximise its competence over all the tasks.

Vieri Giuliano Santucci, Emilio Cartoni, Bruno Castro da Silva et al.

Autonomy is fundamental for artificial agents acting in complex real-world scenarios. The acquisition of many different skills is pivotal to foster versatile autonomous behaviour and thus a main objective for robotics and machine learning. Intrinsic motivations have proven to properly generate a task-agnostic signal to drive the autonomous acquisition of multiple policies in settings requiring the learning of multiple tasks. However, in real world scenarios tasks may be interdependent so that some of them may constitute the precondition for learning other ones. Despite different strategies have been used to tackle the acquisition of interdependent/hierarchical tasks, fully autonomous open-ended learning in these scenarios is still an open question. Building on previous research within the framework of intrinsically-motivated open-ended learning, we propose an architecture for robot control that tackles this problem from the point of view of decision making, i.e. treating the selection of tasks as a Markov Decision Process where the system selects the policies to be trained in order to maximise its competence over all the tasks. The system is then tested with a humanoid robot solving interdependent multiple reaching tasks.

Emilio Cartoni, Gianluca Baldassarre

A parameterized skill is a mapping from multiple goals/task parameters to the policy parameters to accomplish them. Existing works in the literature show how a parameterized skill can be learned given a task space that defines all the possible achievable goals. In this work, we focus on tasks defined in terms of final states (goals), and we face on the challenge where the agent aims to autonomously acquire a parameterized skill to manipulate an initially unknown environment. In this case, the task space is not known a priori and the agent has to autonomously discover it. The agent may posit as a task space its whole sensory space (i.e. the space of all possible sensor readings) as the achievable goals will certainly be a subset of this space. However, the space of achievable goals may be a very tiny subspace in relation to the whole sensory space, thus directly using the sensor space as task space exposes the agent to the curse of dimensionality and makes existing autonomous skill acquisition algorithms inefficient. In this work we present an algorithm that actively discovers the manifold of the achievable goals within the sensor space. We validate the algorithm by employing it in multiple different simulated scenarios where the agent actions achieve different types of goals: moving a redundant arm, pushing an object, and changing the color of an object.