Ontology-Guided Deep Reinforcement Learning for Robotic Task Execution
Sukriti Bhattacharya, Jana Al Haj Ali, Yannick Naudet, Hervé Panetto · IFAC-PapersOnLine · 2025
This paper explores the combination of ontological reasoning and Deep Reinforcement Learning (DRL) as an approach for automated robot capability learning. While ontologies enable robots to autonomously choose and perform tasks through environmental reasoning, they cannot execute unprogrammed actions. Complementarily, DRL cannot be used to reason and make decisions, and it requires predefined objectives and constraints, but can support new behaviour learning to be reprogrammed into the robot. We propose an ontology-guided DRL framework that automatically initializes learning parameters when robots encounter unfamiliar tasks. Using an ontology for capability inference, the system maps ontological knowledge to DRL components including state space, action space, and reward functions. We illustrate with a collaborative assembly scenario, where a robot is not pre-programmed to screw. The ontology is used to infer the feasibility of the task and the required capabilities, while the DRL component is dynamically configured based on this information, allowing the robot to learn the skill autonomously.