The HRIA project strengthened research and innovation capacity in the field of artificial intelligence through the development of methods, technologies, datasets, demonstrators, and prototypes with applications in high-impact areas such as aerial and satellite imagery analysis, energy, medicine, robotics, embedded systems, communications, and the design of materials and integrated circuits.
In the field of perception and understanding of the environment from an aerial perspective, advanced datasets and methods have been developed for 3D reconstruction, depth estimation, and semantic segmentation. Key results include the multimodal ClaraVid dataset, with nearly 17,000 high-resolution frames, as well as UAVid++, which improves the quality and complexity of benchmarks for aerial image segmentation.
In the field of space surveillance, the technologies developed within HRIA were validated in ESA campaigns for monitoring satellite re-entries into the atmosphere. The software solutions for detection, astrometric calibration, and tracking enabled the generation of 105 tracklets during the Rumba, Samba, and Tango campaigns, while the 2026 observations achieved a 100% acceptance rate from ESA.
In the area of energy and smart cities, HRIA results contributed to the development of forecasting and energy consumption management solutions, including systems based on explainable AI and the fusion of data from multiple sensors. These solutions enabled reductions in forecasting errors of up to 29% and peak demand of up to 36.36%. At the same time, the expertise gained contributed to attracting new Horizon Europe projects and strengthening collaborations with industrial and European partners.
In artificial intelligence for medicine, foundation models for OCT imaging, robust methods for detecting mislabeled data, medical image analysis solutions, and diagnostic assistance systems have been developed. The DINOv2-based OCT model achieved a macro-F1 score of 0.775, compared with 0.662 for the generic DINOv2 model, while the OphSurgeon framework improved surgical phase recognition performance by approximately 13.2%.
Important results were also achieved in medical image processing, including the segmentation of anatomical structures and AI-assisted diagnosis. An enhanced MIST architecture incorporating attention mechanisms and Mamba modules achieved a Mean DSC of 90.44% on the ACDC test set for the segmentation of the ventricles and myocardium.
În domeniul roboticii medicale, au fost realizate sisteme pentru navigație robotică în implantologia dentară, roboți chirurgicali, platforme de reabilitare și soluții bazate pe Digital Twin și AI pentru chirurgie minim invazivă. Rezultatele au fost validate experimental pe prototipuri și modele anatomice, demonstrând fezabilitatea integrării AI, roboticii și procesării în timp real.
In autonomous robotics and intelligent control, methods for resilient drone control under cyberattacks, perception and navigation algorithms, and solutions for agricultural robotics were developed. An important research direction focused on aerial object detection, achieving state-of-the-art performance and applicability in agricultural crop analysis.
The project also generated results in natural language and speech processing, including methods for fact-checking in Romanian, neuro-symbolic reasoning, speech synthesis, emotion recognition, and assistive communication technologies based on eye tracking. Several modern text-to-speech synthesis architectures were also developed and evaluated on Romanian datasets.
In the field of materials and integrated circuits, AI-based solutions were developed for alloy design and analog circuit optimization. The AlloyGraph platform combined knowledge graphs, physics-informed machine learning, and multi-agent architectures, achieving 91.2% accuracy on domain-specific questions. For an LDO regulator designed for automotive applications, AI-assisted optimization enabled the maximum output current to be doubled, from 100 mA to 200 mA, without increasing the quiescent current.
Overall, HRIA resulted in public datasets, software technologies, functional demonstrators, and experimental prototypes, supported by international scientific publications and collaborations with industrial and institutional partners. The results demonstrate the ability to transform fundamental research into experimentally validated AI solutions with potential for technology transfer and use in real-world applications.