Safe AI

Safe Autonomy

Overview

Focuses on ensuring the development of trustworthy AI systems through key research areas: AI model robustness verification and improvement; explainable AI (XAI), aimed at improving the transparency of AI decisions; AI reliability assessment, evaluating consistent performance; and safety analysis and assurance for learning-enabled systems, providing systematic and structured safety justifications.

Our work lays the groundwork for integrating AI technologies responsibly into the automotive domain and general society.

Impacts

Our group presented our latest system-agnostic AI safety framework at the United Nations Economic Commission for Europe 23rd GRVA session.

This Operational Design Domain (ODD)-based AI Safety In autonomouS Systems (OASISS) frameworkLink opens in a new window provides practical methods and scientific evidence that AI developers, policymakers, and regulators can incorporate in their frameworks and quantifiably determine if a self-driving system or any AI-powered autonomous system is safe enough for deployment.

Safe AI projects

Explore our key Safe AI research projects to ensure AI technology is developed safely and responsibly.

AIGGREGATE

AIGGREGATE: AI-enhanced collective intelligence for resilient, ethical and user-centric awareness and decision making in CCAM applications

Funding value: €5M, funded by EU Horizon, 500K to WMG
Partners: Eindhoven University of Technology; Vicomtech; TNO; KU Leuven; RWTH Aachen; IDIADA; CERTH; INFINEON; MAPtm; Continental; PAVE

Context and Challenges:
Road safety remains a critical issue in Europe, with over 20,000 fatalities annually. While automated driving offers potential safety, mobility, and efficiency gains, current systems lack the human-like comprehension, predictive capability, and collective decision-making needed in complex, dynamic traffic. Existing CCAM technologies struggle to integrate heterogeneous data, anticipate behaviours, and coordinate ethically across vehicles and infrastructure. Overcoming these limitations requires resilient collective awareness, predictive state modelling, and ethically grounded hybrid intelligence for real-time, human-like decision-making in mixed road environments.

Objectives:

AIGGREGATE aims to transform CCAM from reactive, ego-vehicle approaches to proactive, collective, and socially robust systems. The project’s seven objectives are:

Develop a user-centric ethical framework enabling human-like control.
Create resilient collective awareness models fusing multi-source data into enriched Local Dynamic Maps.
Build predictive state awareness models to anticipate behaviours and system capabilities.
Design hybrid intelligence algorithms for safe, ethical, and human-like decision-making at both vehicle and mobility system levels.
Establish a robust safety assurance framework and tools.
Develop an open-source simulation environment for integration, validation, and assurance.
Demonstrate the integrated system in complex urban and highway scenarios.
By combining AI, human cognition, and ethical governance, the project will deliver CCAM solutions that enhance safety, trust, and traffic efficiency, supporting the broad adoption of connected and automated mobility in real-world environments.

Approach and Innovation:

AIGGREGATE integrates four pillars: 1) an ethics and human factors framework, 2) resilient collective awareness via early-fusion AI models combining vehicle, infrastructure, and user data, 3) predictive state awareness using AI models enhanced with symbolic reasoning and V2X data, and 4) hybrid intelligence for collaborative decision-making. The approach shifts from ego-centric to collective, real-time understanding, enabling vehicles to negotiate manoeuvres, prevent gridlocks, and operate safely beyond current Operational Design Domains. Novel elements include semantic Local Dynamic Maps, capability prediction modules, and ethically embedded decision algorithms. Development will follow an iterative alpha–beta cycle, validated through 2,000+ simulated scenarios and physical demonstrations with multiple OEM vehicles. The open-source simulation environment and safety assurance toolkit will set new standards for testing and validation, supporting transparency, interoperability, and industry uptake.

Impact and Next Steps

AIGGREGATE will advance CCAM safety, efficiency, and societal acceptance through resilient, ethical, and human-like automated driving capabilities. Its innovations, i.e., open simulation tools, safety assurance frameworks, and interoperable decision-making algorithms, will provide lasting resources for researchers, OEMs, and regulators. Demonstrations in complex real-world settings will prove feasibility and build public trust. Long-term, the project’s methods will enable wider deployment of CCAM, integration with traffic management systems, and adaptability to future mobility services. By embedding ethical principles and inclusivity from the outset, AIGGREGATE will guide the responsible evolution of AI-driven mobility well beyond the project’s lifecycle.

Harnessing Synthetic Data Fidelity for Assured Perception of Autonomous Vehicles

Funding value: £500K, funded by UK EPSRC, 500K to WMG
Partners: Denso, Siemens, Adelard (part of NCC group), Wayve UK, Chalmers University of Technology, University of Liverpool, City St George University of London

Project Context and Challenge:

As autonomous vehicles (AVs) move from controlled environments to public roads, safety assurance becomes paramount. Scenario-based testing using synthetic data is central to evaluating AV perception, but there is no standardised method for quantifying synthetic data fidelity. Current approaches focus on realism rather than how machine learning models process and respond to simulated inputs. This gap undermines trust in virtual testing. The project tackles this by defining, quantifying, and using fidelity to improve verification and assurance of AV perception systems.

Objectives:

The project’s primary aim is to integrate fidelity-aware methodologies into the safety assurance of AV perception. Four key objectives guide the work:

1.Fidelity Specification – Define a spectrum of fidelity concepts and associated metrics tailored to AV perception models.

2.Fidelity Quantification – Develop a Safe AI toolchain to train robust predictors that estimate fidelity scores for synthetic data points, calibrated via real-world counterparts.

3.Fidelity-Aware Verification – Enhance debug and acceptance testing by incorporating fidelity metrics, enabling more targeted bug detection and reliability assessment.

4.Validation & Communication – Conduct case studies with industrial partners to validate the methods and create assurance case templates linking fidelity-aware evidence to safety claims.

By aligning fidelity assessment with machine learning behaviour rather than generic realism, the project aims to establish rigorous, practical frameworks for using synthetic data in AV verification and validation, supporting regulatory and industrial adoption.

Approach and Innovation:

The project introduces a novel fidelity-aware safety assurance framework for AV perception. First, it formalises multiple fidelity definitions—covering ML outputs, learned features, and explainability profiles—paired with corresponding metrics. These feed into a fidelity quantification pipeline using Safe AI methods, including adversarial training, robustness testing, uncertainty quantification, and GAN-inspired simulator calibration. The fidelity scores are then integrated into two verification streams: debug testing (prioritising high-fidelity synthetic seeds for bug discovery) and acceptance testing (modelling the statistical impact of fidelity on reliability claims). The approach is validated through case studies using industrial-grade simulators, generative AI models, public datasets, and the SafetyPool™ scenario database. Assurance case templates are developed to communicate results within regulatory frameworks. This is the first comprehensive methodology to operationalise fidelity-aware synthetic data in AV testing, shifting from visual realism toward ML-behavioural alignment, thus bridging simulation and real-world safety evidence.

Impact and Next Steps:

The project will deliver the first standardised fidelity specification, quantification tools, and assurance frameworks for synthetic data in AV perception testing. Its outputs—open-source predictors, verification methods, and safety case templates—will directly benefit industry partners and feed into UK AV safety standards, supporting CAVPASS and related initiatives. By enabling safer, faster, and more cost-effective AV development, the work accelerates deployment while strengthening public trust. Beyond AVs, the fidelity-aware methodology can be adapted to other domains reliant on virtual testing, such as robotics, aerospace, and nuclear systems, establishing a cross-domain benchmark for trustworthy synthetic data use in safety-critical AI applications.

What we can offer

Underpinned by scientific evidence, our research helps you put the users at the heart of your technology developments and deploy innovations in the real world at the highest safety standards.