mUlti-Level Trustworthiness to IMprove the Adoption of hybrid arTificial intelligencE

Project Information

ULTIMATE

Grant agreement ID: 101070162

DOI

10.3030/101070162

Project closed

EC signature date 12 June 2022

Start date 1 October 2022

End date 30 September 2025

Funded under

• Digital, Industry and Space

Total cost

€ 3 529 112,50

EU contribution

€ 3 529 112,50

Coordinated by

THALES
France

Periodic Reporting for period 2 - ULTIMATE (mUlti-Level Trustworthiness to IMprove the Adoption of hybrid arTificial intelligencE)

Reporting period: 2024-04-01 to 2025-09-30

Summary of the context and overall objectives of the project

Artificial Intelligence (AI) is increasingly used in business, offering productivity and innovation gains, but faces adoption barriers in industry. Hybrid AI (HAI), which combines data-driven and physics-based approaches, shows promise but still struggles with trustworthiness, including validity, explainability, and ethics. The ULTIMATE project addressed these challenges by developing advanced data representation and visualization models for HAI, enabling clear, human-understandable insights. It introduced new architectural patterns and training strategies to integrate data-driven, symbolic, and physics-based components in a trustworthy way. The project also created evaluation methodologies tailored to HAI, focusing on accuracy, generalization, interpretability, and resilience. These advances were tested in real industrial environments, demonstrating technical and operational feasibility. ULTIMATE ensured compliance with ethical, legal, and societal standards, and promoted adoption through engagement at European and international levels, supporting a sustainable ecosystem for trustworthy HAI in industry.

Work performed from the beginning of the project to the end of the period covered by the report and main results achieved so far

Building on a set of specific KPIs (comprising quantitative and qualitative measures to evaluate efficiency and effectiveness in real-world industrial settings) tailored to three representative Use Cases (anomaly prognosis on satellites, human–robot collaboration in a robotic workshop, and object recognition and autonomous navigation in logistic chains), ULTIMATE has systematically addressed the Design & Development, Evaluation, and Execution under operational conditions stages of the HAI life cycle to ensure AI trustworthiness. Across these stages, the project has achieved significant progress in implementing HAI algorithms in industrial environments and in assessing their performance, robustness and usability in practice, thereby demonstrating the feasibility and added value of trustworthy hybrid AI solutions in diverse operational contexts.

ULTIMATE has elaborated a data visualisation toolkit for representing high-dimensional information. The use of visualisation toolkit for monitoring the network evolution and the training of deep-learning models also offers insight into the quality of the training data.

The project has developed advanced HAI algorithms tailored to specific industrial UC:
- an artificial neuronal network approach, completed by a rules-based AI approach for reasoning and explainability, to earlier detect and classify anomalies on satellites to alleviate the human burden at ground
- a neuro-symbolic solution that combines Deep Learning (DL) with symbolic knowledge for reasoning in front of potential safety issues due to a shared space between humans and robots (visual tracking, event recognition, and logical reasoning)
- a DL considering distance estimation (object, recognition, scene understanding) and a combination of RL / Markov decision processes which aim at coordinating robots that sort and bring (planning and navigation) different types of objects within an industrial logistic chain.

ULTIMATE evaluated the HAI algorithms through the implementation or the adaptation of several tooled-up methods addressing the chosen trustworthy AI criteria (reliability, robustness, explainability), and a rigorous iterative process, which enhanced the initial HAI algorithms. A structured evaluation process was established to test the trustworthiness of HAI. Tools were created to measure how systems respond to uncertainty, unexpected inputs, and changing conditions. Simulation environments and statistical tests helped ensure the AI models behaved consistently and could explain their decisions. Feedback between developers and evaluators improved both the algorithms and the testing tools.

The HAI solutions were integrated into the three UC. Each platform tested how HAI performs in realistic environments. Results showed improved safety, reliability and efficiency compared to conventional AI. Stakeholder feedback confirmed that transparent explanations and cautious system behaviour increased user trust.

Across all areas, the aim was to make AI systems more transparent and adaptable while maintaining human oversight. An ethical and legal review ensured compliance with data protection, fairness, and transparency principles. Ethical reviews verified that privacy, accountability, and fairness requirements were met. A resulting end-to-end AI trustworthiness approach has been experimented through the three UC to consider both technical and ethical aspects at the same level.

Progress beyond the state of the art and expected potential impact (including the socio-economic impact and the wider societal implications of the project so far)

ULTIMATE

- built a toolkit to display complex information (raw data) in ways humans can easily interpret. This toolkit allows the teams to visualise large datasets and understand patterns, spot irregularities, and track performance more effectively. It became a foundation for the rest of the project.

- created new ways to build and train HAI solutions that people can trust. For space monitoring, the AI approach can spot different kinds of anomalies and clearly explain why it makes certain decisions. For workshop safety, the HAI solution helps robots understand human behaviour and act safely, using transparent rules that humans can follow and adjust. In logistic chains, DL/RL techniques are combined with others so robots can detect people and objects, move safely, and perform precise tasks. These innovations show practical, trustworthy uses of HAI.

- developed structured (statistical, formal and empirical) methods to evaluate how well the HAI algorithms behave. They used tools to measure three key qualities: how reliable the algorithms are, how robust they are to errors, and how clearly they can explain their decisions. These evaluations were done iteratively to improve the algorithms over time. Alongside the technical checks, ULTIMATE also assessed possible ethical and legal risks. By combining numerical measures with qualitative judgments, a balanced view of performance and responsibility has been created.

- tested their hybrid AI system directly in industrial settings to show that it works reliably in real operations. The system combines smart perception, such as recognizing objects, people, and situations, with decision-making tools that help robots choose and perform the right actions safely and efficiently. It can also detect unusual situations and ignore predictions it does not trust.

- built ethics and trust into every step of the project. Dedicated experts oversaw data protection, legal compliance and responsible use. They regularly updated a data-management plan to follow European rules. Stakeholders’ values were gathered and turned into concrete system requirements. Technical teams then created tools to make the AI reliable, safe, understandable and respectful of privacy. Independent reviews checked that the evidence matched legal and ethical duties. Finally, real-world demonstrations showed these safeguards working in practice, including human oversight, safe-stop features, transparency, logging and privacy-preserving data handling.

The project proved that responsible AI can drive innovation, public trust, and high-quality human-AI collaboration.