- Home
- About TUAI
- Research
- Doctoral Candidates
- Networking
- Consortium
- Contact
WP3 – Research Progress 2026
WP3 aims to combine data from multiple sensors and intelligent devices to improve perception, decision-making and overall system performance in industrial environments. The work addresses the challenges of integrating heterogeneous sensor data from sources such as cameras, LiDAR, radar, ultrasonic sensors and GPS, with particular emphasis on cross-modal alignment, synchronisation and efficient fusion. It also focuses on developing accurate object and obstacle recognition models and lightweight AI solutions capable of real-time operation on autonomous guided vehicles (AGVs), drones and industrial robots. An important objective is to translate advances in sensor fusion and machine learning into practical industrial applications, ensuring that developed methods can operate efficiently and reliably under real-world conditions.
During the Summer School, doctoral candidates involved in WP3 presented their individual research projects in the context of these shared objectives, highlighting complementary research directions within the Work Package. Their presentations addressed key challenges in industrial perception, including the integration and interpretation of heterogeneous sensor data, object and obstacle recognition, and the development of efficient AI models for resource-constrained and real-time applications. Together, the projects demonstrated how advances in sensor fusion and machine learning can support more reliable perception and decision-making in industrial environments, particularly in applications involving AGVs and other autonomous robotic systems.
Pi-Wei Chen (SUT) presented his research on extending collaborative sensor fusion towards autonomous industrial systems, focusing on how AI-based perception can be connected with reasoning and autonomous decision-making. His doctoral project, Enhancing Perception Accuracy through Collaborative Sensor Fusion Framework in Autonomous Systems, aims to develop systems that use fused sensor data for prediction tasks, including anomaly detection and condition-based maintenance, and subsequently use these predictions to determine appropriate actions. His presentation explored how agentic AI could support this transition from perception to autonomy.
Pi Wei introduced the concept of agentic retrieval-augmented generation (Agentic RAG), highlighting the limitations of conventional RAG systems, which retrieve information primarily according to similarity, and knowledge-graph-based approaches with predefined retrieval strategies. He formulated a research problem concerning the optimisation of retrieval policies: how an agent can select the most appropriate search, expansion and reading actions based on the evidence already collected and the information still missing. The objective is to improve both the accuracy of evidence-grounded answers and retrieval efficiency.
He then presented a proposed prototype framework in which a frozen large language model (LLM) is combined with an external retrieval skill that can be improved through experience without modifying the model’s weights. The framework uses a unified retrieval substrate containing documents, chunks, sentences and entities, enabling multi-level retrieval and graph-like navigation across related information. At query time, the agent selects among four actions—SEARCH, EXPAND, READ and FINISH—and updates its evidence state after each retrieval step. The proposed approach aims to improve evidence acquisition and decision-making while retaining the underlying LLM and retrieval substrate unchanged. The presentation outlined the research question, methodological rationale and prototype architecture; it did not report a completed experimental validation of the proposed framework.
Alexandre Niyomugaba (SUT) presented his research on GPU-accelerated edge computing and federated learning for industrial environments, focusing on efficient AI processing on resource-constrained devices. His work contributes to WP3 by addressing the deployment of AI models in industrial systems, where real-time performance, limited computational resources and energy efficiency are important requirements. His research combines federated learning, edge computing and GPU acceleration to support distributed anomaly detection, fault prediction and intelligent decision-making close to the source of sensor data.
He presented progress on Edge-FCL, a federated continual learning framework for energy-efficient anomaly detection at the industrial edge. The approach addresses the challenges of detecting anomalies in non-stationary visual data under conditions of limited local data and heterogeneous data distributions. The experimental setup uses the MVTec AD dataset, comprising 5,354 images across 15 categories, and a multi-device edge architecture. His evaluation highlighted the trade-off between prediction performance and resource consumption: running the model exclusively in FP32 precision increased average inference time from 28.2 ms to 47.3 ms and memory consumption to 391.05 MB, while achieving 89.44% accuracy. These results illustrate the importance of adapting AI computation to the constraints of industrial edge devices.
Alexandre also presented research on “Edge-Based Safety Optimization: Lightweight Fuzzy Integrated Model for Trustworthy AGV Braking.” This work addresses the limitations of conventional rule-based braking, including abrupt stops, mechanical wear and disruption of industrial logistics. The proposed framework combines a lightweight 2D-CNN–GRU perception model with a fuzzy logic controller to select between continuing, slowing down and stopping according to obstacle distance, time to collision and estimated motion intensity. The presentation reported a model with 814,595 parameters, a mean squared error of 0.00245 and an (R^2) prediction fidelity of 89.15%. The framework aims to provide interpretable and adjustable braking decisions while reducing mechanical stress on AGVs. Alexandre also identified collaborative perception of AGV trajectories as a direction for future research.
GPU-Accelerated_Edge_Computing_for_Federated_Learning_Reasoning_in_Industrial_Environments.pdf
Myroslav Mishchuk (SUT) presented his research on explainable artificial intelligence for internal logistics systems in green manufacturing, proposing a shift from explanations that merely describe model predictions towards explanations that can support system adaptation and control. His work contributes to WP3 by addressing the reliability and interpretability of AI in industrial perception and autonomous systems, particularly where sensor fusion and real-time decisions affect the safe operation of autonomous guided vehicles (AGVs). He argued that, in safety-critical industrial environments, explanations should not only help users understand AI outputs but also provide a basis for detecting weaknesses and improving system behaviour.
He reviewed his previous work on explainability for individual sensor-based predictions and industrial AI applications, including LiDAR-based collision avoidance, Wi-Fi signal-strength prediction and battery remaining-useful-life estimation. These studies provide a foundation for his current investigation into the reliability of temporal explanations. He identified a key limitation of post-hoc explainability methods such as TimeSHAP: the use of unconditional background values can produce unrealistic masked sequences, potentially distorting feature attributions when time-series data exhibit non-stationary trends. To address this issue, he proposed Trend-Aware Backgrounds (TAB), including state-conditional and past-trajectory-based approaches designed to produce more appropriate reference values while preserving temporal information and avoiding data leakage.
Myroslav also presented a broader conceptual framework for “explanation-as-control,” organised around four conditions for translating explanations into action: faithfulness, the level at which an explanation is provided, agency, and the ability to demonstrate that the resulting feedback loop is stable and beneficial. He outlined a benchmark combining calibrated synthetic time series with real battery-life datasets to investigate attribution bias and evaluate the proposed methods. At the time of the presentation, the methodology had been finalised and the experimental evaluation was about to begin. As a longer-term research direction, he discussed using explanations to guide adaptive learning and reconfiguration in industrial AI systems, including AGV navigation and braking, while emphasising the need to validate the effects of any resulting changes.
Explainable_Artificial_Intelligence_for_internal_logistics_systems_in_green_manufacturing.pdf