Trustworthy AI for CCAM

Informazioni relative al progetto

AI4CCAM

ID dell’accordo di sovvenzione: 101076911

DOI

10.3030/101076911

Progetto chiuso

Data della firma CE 21 Ottobre 2022

Data di avvio 1 Gennaio 2023

Data di completamento 31 Marzo 2026

Finanziato da

• Climate, Energy and Mobility

Costo totale

€ 5 965 630,00

Contributo UE

€ 5 965 630,00

Coordinato da

SIMULA RESEARCH LABORATORY AS
Norway

Periodic Reporting for period 1 - AI4CCAM (Trustworthy AI for CCAM)

Periodo di rendicontazione: 2023-01-01 al 2024-06-30

Sintesi del contesto e degli obiettivi generali del progetto

Considering AI capabilities and potential risks, AI4CCAM develops an open environment for integrating trustworthy-by-design AI models of vulnerable road user behaviour anticipation in urban traffic conditions and accounting for improved road safety and user acceptance. Leveraging the Trustworthy AI guidelines for general systems and the ethics recommendations for connected automated vehicles (AVs), AI4CCAM supports AI-based scenarios management in which road user behaviour anticipation models integrate visual gaze estimation and where explainable trajectory prediction models are simulated and multiplied with generative adversarial networks and metamorphic testing. The AI4CCAM open environment includes an interoperable digital framework and an online participatory space to foster acceptance of AI in AVs. Simulation scenarios are developed in three complementary use cases covering the whole sense-plan-act paradigm and user acceptance. As such, AI4CCAM advances knowledge in building trustworthy-by-design AI-based solutions for CCAM applications. SSH research methods are used in AI4CCAM on user acceptance of AVs and in the AI4CCAM participatory space where participative democracy is used.

Lavoro eseguito dall’inizio del progetto fino alla fine del periodo coperto dalla relazione e principali risultati finora ottenuti

The development of the AI4CCAM interoperable digital framework is ongoing and its architecture has been specified during RP1. As part of this digital framework, the AI4CCAM Trustworthy AI methodology has been delivered (D1.1) 12 urban traffic scenarios (functional and logical) have been modelled in MOSAR, and one test protocol is ready for simulation. An initial version of the digital framework has been developed based on the provided results, while its final version will be completed during RP2. Based on these elements R1.1 and R1.3 are completed and R1.2 (second version of the digital framework) is expected at the end of the project.

Trustworthy AI for CCAM through ethical, social and cultural implication is defined and activities are ongoing to improve user acceptance of automated vehicles. The AI4CCAM participatory space (R2.1) is up and running and now expected to be populated by EU citizen contributions. It presents a knowledge base related to AI risks, ethical risks in CCAM and cyber threats for automated vehicles (R2.2). The Ethical and Scientific Advisory Board has been setted up and already met 3 times (R2.3). The production of a roadmap for user acceptance of automated vehicles by VRU is ongoing (R2.4).

Three use cases for exhibiting VRU-AV interactions with potential ethical risks are ongoing. The evaluation criteria related to ethics have been set and are available as part of the AI4CCAM validation handbook (R3.1 D4.1). The CAV-VRU interaction demonstrator using Virtual Reality is well ongoing (R3.2 D4.2). The design of simulation scenarios for the three use cases is ongoing (D4.4). The implementation of the three use cases has started and we are confident that they will be ready to evaluate the multiple research results obtained in the project (R3.3 D4.3).

Research results have been reached with the preparation of multiple scientific articles, that are either published according to open-science practices (at least 6), accepted or submitted for publication (R4.2). Trustworthy-by-design models are being trained or validated with GANs and Metamorphic Testing. These models are designed for pedestrian movement anticipation, scene understanding, and visual gaze estimation (R4.1). Regarding education already two PhD students have been educated on Trustworthy AI for CCAM (R4.3). At mid-term of the project, objective SO4 is well on track and will be reached at the end of the project.

Regarding outreach, many communication and dissemination activities took place during RP1. Apart from the delayed Stakeholders Forum (R5.1°, which was organised in May 2024, instead of Fall 2023, all the planned C&D and exploitation actions have been performed, including a market analysis and first exploitation roadmap (R5.2) and innovation registry (R5.3).

Progressi oltre lo stato dell’arte e potenziale impatto previsto (incluso l’impatto socioeconomico e le implicazioni sociali più ampie del progetto fino ad ora)

During RP1, the AI4CCAM project has advanced the SotA in multiple areas including:
1. The development of trustworthy-by-design AI models by proposing an abstract representation of sensor-based urban scenes including vulnerable road users;
2. The identification of ethical risks in scenarios and datasets by proposing indicators related to the Trustworthy AI ethical guidelines;
3. The user acceptance of automated vehicles by performing SSH research;
4. The design and development of simulation scenarios for testing AI models;
5. The development and test of advanced human trajectory prediction models;
6. The development of gaze prediction models;
7. The detection and mitigation of cyber threats in CCAM;

These advances have resulted in the production of 12 scientific articles, 6 being already published in open-access.

AI4CCAM Project