Abstract
This paper presents a semantic framework designed to enhance the autonomy of robots operating in complex and irregular environments. By integrating human expert knowledge through a multi-layered ontology and employing a hierarchical mission control architecture, the system enables robots to interpret sensor data more effectively. The framework utilizes Bayesian networks for decision-making, allowing robots to assess the confidence of their findings and autonomously initiate corrective actions, such as adjusting perspectives, to resolve environmental uncertainties.
Introduction
While robotic hardware and sensor data analysis have seen significant advancements, achieving full autonomy in inspection tasks remains a challenge. Current systems often excel at data collection but lack the capacity to interpret findings or make high-level decisions without human intervention. This is particularly problematic in complex environments like sewers or rugged outdoor terrains, where a robot may encounter features of interest—such as defects or waste—but lack the confidence to classify them accurately. This research addresses the gap between raw data processing and high-level mission planning, proposing a system that can plan, execute, and adapt based on a semantic understanding of its surroundings.
Methodology
The proposed system relies on a multi-layered approach centered on semantic techniques to bridge the gap between low-level control and high-level reasoning.
Hierarchical Architecture
The mission control system is organized into four distinct levels, ranging from the "Subsymbolic Level" (handling raw sensor and actuator interfaces) to the "Semantic Level" (responsible for high-level management and goal decomposition). This structure allows for a clear separation between hardware-specific tasks and abstract mission logic.
Ontology-Based Knowledge Base
A central component of the framework is the Knowledge Base, implemented using OWL-DL and the KAON2 framework. Expert knowledge is organized into three hierarchical layers:
- Basic Ontology: Covers fundamental concepts such as time and parameters.
- Core Ontology: Defines general robotics concepts including robot types, environments, and navigation.
- Domain Ontology: Contains application-specific knowledge, such as details about the "Hexapod Walking Machine" or specific waste detection criteria.
Decision-Making and Execution
Mission plans are modeled as "Flexible Programs"—tree structures consisting of branching, action, and planning nodes. To handle uncertainty, the system employs Bayesian networks for temporal data fusion, combining multiple observations of an object over time. Bayesian decision networks are then used to determine whether the robot should continue inspecting an object to increase confidence or proceed with the mission. The entire system is implemented using the MCA2 modular C++ framework.
Experimental Setup
The framework was evaluated using a custom simulation environment rather than standard public datasets.
- Robot Model: The simulation utilized a digital twin of the LAURON IVc, a six-legged (hexapod) walking machine designed for rugged terrain.
- Scenario: The test cases involved the detection and classification of various types of waste in diverse irregular terrains, including river banks, dunes, and forests.
- Validation: Component stubs were used to simulate sensor data, allowing researchers to focus on validating the logic of the planning and decision-making components.
Discussion
A key innovation of this system is its "self-directed curiosity." Instead of following a rigid path, the robot actively examines entities of interest. If initial data analysis yields low confidence, the semantic model triggers subgoals, such as approaching the object from a different angle. Furthermore, the explicit separation of knowledge representation from execution control makes the system highly extensible; adding new inspection targets requires only an update to the ontology rather than a complete overhaul of the navigation software. This semantic approach also facilitates better human-robot communication, as users can interact with the system using abstract goals.
Results and Future Work
Preliminary results focus on the functional validation of the architecture. The system successfully demonstrated the ability to decompose high-level goals into executable programs and handle simulated hardware errors through automated replanning. Simulation tests confirmed that the Bayesian logic correctly prioritized mission objectives based on real-time confidence levels.
Future work will focus on transitioning from simulation to real-world field tests with the LAURON hexapod to address challenges like sensor noise and physical manipulation on irregular terrain. Additionally, the authors aim to develop a more advanced semantic user interface and incorporate machine learning to optimize resource usage over time.
Summary updated