Investigating Knowledge Gain through Ontological Epistemics for Autonomous Robot Exploration

Investigating Knowledge Gain through Ontological Epistemics for Autonomous Robot Exploration 1st Gloria Beraldo CNR-ISTC Italy [email protected] 2nd Angelo Oddi CNR-ISTC Italy [email protected] 3rd Riccardo Rasconi CNR-ISTC Italy [email protected] 4th Andrea Orlandini CNR-ISTC Italy [email protected] 5th Alessandro Umbrico CNR-ISTC Italy [email protected] Abstract—Humans acquire knowledge through dynamic interactions with their environment, especially when their actions are guided by underlying semantics. This study presents a cognitive framework designed to replicate such human-like capabilities within a robotic system, alternating between exploration and exploitation. Our primary focus is on structuring the knowledge acquisition process: the framework is grounded in an epistemic model that captures and organizes relevant information in a structured, semantically rich, and context-aware manner—enabling continuous and adaptive learning over time. Index Terms—Knowledge Representation and Reasoning, Ontology, Cognitive Robotics I. INTRODUCTION Robotic agents must develop adaptive deliberative capabilities to operate autonomously in unstructured environments without a priori knowledge of the domain context [1]. Traditional approaches for promoting robot autonomy fall into two main categories [2]: (a) AI-driven methods adopt a topdown perspective, often abstracting from low-level control and physical constraints; (b) Robotics-driven methods follow a bottom-up approach, focusing on motion and control while overlooking high-level reasoning and action composition. These perspectives are complementary and both essential for achieving robust autonomy. A key challenge lies in enabling robots to incrementally build autonomy by grounding action knowledge from physical and geometric interactions with the environment. We address the challenge of enhancing perception by equipping robots with semantics models to interpret the outcomes of their actions and incrementally build knowledge (especially when they lack a priori knowledge of the environment beyond primitive skills). Research efforts have explored data-driven approaches that abstract sensorimotor data into action models for goal-directed planning [3], [4]. While promising, these methods often lack semantic grounding of clustered states and conditions, limiting both generalization across tasks and environments, and preventing systematic control over the properties involved. In this regard, we are investigating a different approach relying on an ontological epistemic model of robot-interacting skills. Coherently with other research [5]–[7], we investigate the use of foundational This work is partially supported by the Italian Ministry of Research, under the complementary actions to the NRRP “Fit4MedRob - Fit for Medical Robotics” Grant (PNC0000007 - CUP: B53C22006990001) and by PNRR MUR project PE0000013-FAIR. Fig. 1. Cognitive architecture for semantic-based exploration and exploitation. ontologies to enhance the understanding of the world from direct experience. We center our approach on using ontologies to unlock the deeper epistemic meaning of robot skills, viewing each action not just as execution but as a perceptual experience. By doing so, every action reshapes the robot’s internal belief, growing its knowledge of reachable states, success and failure conditions, and the skills it can activate along the way. With this purpose, we design a cognitive architecture, shown in Fig. 1, that empowers a robot to: (a) actively explore its environment, uncovering new entities and their properties (exploration stage), (b) leverage its consolidated knowledge to strategically guide future actions (exploitation). This work outlines key parts of the exploration, guided by an epistemic model for acquiring contextual knowledge. II. FROM PERCEPTION TO KNOWLEDGE ACQUISITION We assume the robot, according to its embodiment, possesses a set of skills it can perform in the environment, such as moving to a position, detecting objects, picking objects, and placing objects. However, the robot initially lacks knowledge of how to combine skills (i.e., planning) to effectively expand its understanding of the environment and create more complex behaviors aimed at reaching high-level goals. Considering the embodiment of a robot (i.e., its structural composition) as known, we encode a small amount of domain-independent knowledge into a semantic layer (see Fig. 1) to support the interpretation of skill execution. This semantic layer characterizes the epistemic information gain consequent to the success or failure of skills. Taking inspiration from active inference [8], we design an Active Perception module that guides the robot in developing an awareness of possible environmental 2025 I-RIM Conference October 17-19, Rome, Italy ISBN: 9788894580570 10.5281/zenodo.17629752 123 interactions (social belief). For example, we expect the skill ObjectPicking(oj)to be triggered only if the object oj has been previously detected and is visible from the robot’s current position. To learn such preconditions from scratch, the robot begins by randomly executing its available skills, which results in a high number of failures, as its initial social belief is minimal. Over time, it gradually develops more intentional and context-aware behaviors based on the physical and relational properties it discovers through interaction. The belief is constructed and refined through the use of an ontology, which ensures the semantic coherence and contextual framing of the acquired knowledge, enabling the robot to reason meaningfully about its actions and their outcomes. Specifically, in our framework, the ontological module is based on DOLCE [9] and extends the concept of Function [10] to capture the epistemic effects of robot skill and action execution, in addition to functional ones. III. EPISTEMIC KNOWLEDGE ACQUISITION To collect feedback and store the discovered relations and properties of the environment, our framework relies on a knowledge graph (experiential graph), that is dynamically generated and updated during the robot’s exploration. It is modeled as a bipartite graph, where one set of vertices represents perceptive states Pand the other corresponds to the specific actions Aassociated with known skills. From the execution of an action a∈ A, two edges are created connecting p−→ aand the a−→ p′, where prepresents the initial state and p′the resulting state after executing a. The arc a−→ p′annotates the outcome of the action (1 success vs. 0 failure). Each vertex is associated with a formal ontological structure, grounded in the epistemic model, which captures the semantics of the entities it denotes. In particular, we adopt general concepts such as PhysicalObject and Quality, which represent measurable aspects whose values are defined within a Region. To incorporate perceptual information, the Belief is modeled as a special kind of Quality attributed to the agent. Meanwhile, the BeliefUpdate is introduced as a SocialObject that encapsulates the outcome of sensory feedback, enabling the system to progressively enrich its understanding of PhysicalObject through repeated interactions and experiential learning. For example, in a delivery scenario where the robot must navigate the environment to pick up and place objects, it is crucial to collect knowledge about which physical locations are accessible, observable, or suitable for manipulation. The property isReachable represents locations known to be reachable by the robot after the successful execution of a GoTo skill. A positive feedback from this skill updates the experiential graph by introducing an individual of ReachabilityBelief, which reifies the observation that the robot’s target configuration is reachable from its starting configuration. Similarly, the property hasObservableLocation formally describes locations that have yielded positive feedback from the ObjectDetection skill. This feedback is reified by an individual of VisibilityBelief, which asserts that a PhysicalObject, along with estimated quality attributes (e.g., location), is visible from the robot’s current configuration. Likewise, the property hasGraspableLocation captures physical locations associated with objects that have been confirmed as graspable, based on a successful execution of the ObjectPicking skill. Starting from this exploratory representation, the bipartite graph is progressively transformed into a Markov Decision Process (MDP) by abstracting perceptive states as MDP states and available actions as MDP actions. The multiple arcs between states and actions are consolidated into single transitions, and the transition matrix is derived by estimating probabilities based on observed successes and failures during execution. This translation from the bipartite knowledge graph to a traditional MDP offers two main advantages: (a) it simplifies the structure and coherently reduces the complexity of the graph, and (b) it enables the estimation of knowledge gain resulting from the execution of a given skill, expressed as a utility function derived from the probabilistic transition matrix. In this way, the system evolves from reactive exploration to goal-directed exploitation, where decisions are informed by the learned structure of the environment and the estimated utility of executing each action in a given state. IV. CONCLUSION In this work, we briefly presented a framework to bridge low-level perceptual feedback with high-level semantic knowledge acquisition through the dynamic construction of a knowledge graph. By leveraging ontological structures, the system grounds perceptual data in meaningful concepts, supporting long-term learning. As a next step, we aim to implement and iteratively refine its components based on empirical evaluation. REFERENCES [1] F. Ingrand and M. Ghallab, “Robotics and artificial intelligence: A perspective on deliberation functions,” 2014. [2] K. Rajan and A. Saffiotti, “Towards a science of integrated AI and Robotics,” 2017. [3] G. Konidaris, L. Pack Kaelbling, and T. Lozano-Perez, “From Skills to Symbols: Learning Symbolic Representations for Abstract High-Level Planning,” tech. rep., 2018. [4] G. Beraldo, A. Oddi, and R. Rasconi, “An Empirical Study of Grounding PPDDL Plans for AI-Driven Robots in Social Environment,” in Frontiers in Artificial Intelligence and Applications, vol. 392, pp. 4426–4433, IOS Press BV, 10 2024. [5] S. Lemaignan, M. Warnier, E. A. Sisbot, A. Clodic, and R. Alami, “Artificial cognition for social human–robot interaction: An implementation,” Artificial Intelligence, vol. 247, 2017. [6] M. Tenorth and M. Beetz, “Representations for robot knowledge in the KnowRob framework,” Artificial Intelligence, vol. 247, pp. 151–169, 2017. [7] L. Ferrini, A. Andriella, R. Ros, and S. Lemaignan, “From Percepts to Semantics: A Multi-modal Saliency Map to Support Social Robots’ Attention,” J. Hum.-Robot Interact., 6 2025. [8] K. Friston, F. Rigoli, D. Ognibene, C. Mathys, T. Fitzgerald, and G. Pezzulo, “Active inference and epistemic value,” Cognitive Neuroscience, vol. 6, no. 4, pp. 187–214, 2015. [9] S. Borgo, R. Ferrario, A. Gangemi, N. Guarino, C. Masolo, D. Porello, E. M. Sanfilippo, and L. Vieu, “DOLCE: A descriptive ontology for linguistic and cognitive engineering,” Applied Ontology, vol. 17, no. 1, pp. 45–69, 2022. [10] S. Borgo, R. Mizoguchi, and B. Smith, “On the ontology of functions,” Applied Ontology, vol. 6, no. 2, pp. 99–104, 2011. 124