Modern robotic systems are evolving toward the capability to model, perceive, plan, and execute tasks based on knowledge about their environment. In such systems, adaptability—namely, the ability to perform assigned roles successfully even in contin...
Modern robotic systems are evolving toward the capability to model, perceive, plan, and execute tasks based on knowledge about their environment. In such systems, adaptability—namely, the ability to perform assigned roles successfully even in continuously changing environments—has emerged as a critical requirement. However, environment-adaptive modeling presents several limitations. The environmental factors to be considered are diverse and complex, and the cost and time required to construct new models for new environments hinder practical deployment. To address these challenges, researchers have proposed advanced approaches that leverage domain knowledge or incorporate large-scale language models to improve the efficiency and accuracy of environmental modeling.
In this dissertation, we propose an integrated framework that enables autonomous environmental semantic extraction and modeling for robots by combining domain knowledge with Large Language Models (LLMs) and Vision-Language Models (VLMs). In the proposed system, a robot extracts semantic information about itself and, once deployed in a new environment, autonomously explores the space to generate a spatial map while simultaneously collecting semantic information about objects and places, which is stored in a structured database. Furthermore, this research goes beyond robot-only environmental modeling and introduces a human-robot collaborative modeling scheme. This allows users to directly participate in the environmental modeling process, thereby enabling effective interaction and knowledge sharing between humans and robots. To achieve this, we define semantic representations for robots, objects, and places, tailored to service robot environments. By specifying the robot’s affordances and the functional properties of objects and places, we enable the robot to understand its environment in a manner similar to humans. To autonomously and training-free collect this semantic information in new environments, robot-related semantics are extracted by querying Unified Robot Description Format (URDF) files and manuals with an LLM; object information is obtained by combining zero-shot object detection and vision question answering with VLMs; and place semantics are inferred using an LLM based on the collected object information. In particular, domain knowledge is incorporated during the extraction process to constrain the target environment's domain, thereby improving the accuracy of semantic extraction. To validate the applicability of the proposed framework to service robots, experiments were conducted in both simulation and real-world environments, and its effectiveness was demonstrated through qualitative and quantitative evaluations. Furthermore, a Robot Serviceability Measurement (RSM) is proposed to assess the framework from a service-oriented perspective, thereby demonstrating its practical utility. Finally, we outline future research directions to further advance autonomous environmental extraction and modeling.