Formal verification for safety evaluation of autonomous vehicles: an interview with Abdelrahman Sayed Sayed

2026-04-14 · Source: ΑΙhub · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Software Development & Engineering · Depth: Expert, medium

Summary

Abdelrahman Sayed Sayed, a Marie Skłodowska-Curie PhD Fellow at Université Gustave Eiffel, is conducting interdisciplinary research on formal verification of neural Ordinary Differential Equations (ODE) for safety evaluation in autonomous vehicles. His work aims to build trust in AI perception models for safety-critical applications by providing formal guarantees on their behavior, addressing concerns from industrial partners and regulatory bodies like the EU. His research comprises three main directions: establishing formal relations and error bounds between discrete and continuous neural models (Neural ODE and Residual Networks) to enable verification proxy; developing new, lightweight interval-based reachability analysis methods for Neural ODE by extending mixed monotonicity approaches; and creating a full Neural ODE verifier toolbox that includes falsification checks and a verification/refinement loop to identify safe and potentially unsafe operational regions. Applications include real-time monitoring in guided railway vehicles and robustness verification for classifiers used in marine autonomous underwater vehicles for zooplankton monitoring.

Key takeaway

For AI Scientists and Machine Learning Engineers developing safety-critical autonomous systems, integrating formal verification for neural ODEs is crucial. Your work should focus on establishing formal guarantees to meet regulatory requirements and build industrial trust, particularly when deploying AI in applications interacting with humans or critical infrastructure. Consider applying lightweight reachability analysis and verification proxies to make these methods computationally feasible for real-world scenarios.

Key insights

Formal verification of neural ODEs can provide critical safety guarantees for AI in autonomous systems.

Principles

• Formal guarantees build trust in AI.
• Lightweight verification methods are crucial.
• Verification proxies reduce computational cost.

Method

A Neural ODE verifier architecture involves initial falsification, followed by a verification and refinement loop that splits input sets to identify safe reachable sets or uncertain regions.

In practice

• Verify AI classifiers in railway systems.
• Assess robustness of marine organism classifiers.

Topics

• Formal Verification
• Neural ODE
• Autonomous Vehicle Safety
• Reachability Analysis
• Marine Robotics

Best for: AI Scientist, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• VASO: Formally Verifiable Self-Evolving Skills for Physical AI Agents — Formal verification can guide LLM-generated robot skill evolution, enhancing trust and compliance.
• Physical AI that Moves the World — Qasar Younis & Peter Ludwig, Applied Intuition — Applied Intuition builds physical AI for safety-critical moving machines, focusing on simulation, operating systems, and fundamental AI models.
• Reinforcement learning applied to autonomous vehicles: an interview with Oliver Chang — Adversarial reinforcement learning can uncover critical vulnerabilities in autonomous cyber-physical systems.
• TNODEV: Toolbox for Neural ODE Verification — TNODEV is the first sound formal verifier for neural ODEs, integrating advanced techniques for precise, iterative verification.
• On the Formal Limits of Alignment Verification — AI alignment verification faces a trilemma: soundness, generality, and tractability cannot be simultaneously achieved.
• Formal Methods as Agent Guardrails — Neuro-symbolic AI combines formal methods and LLMs to define and verify agent behavior, dramatically scaling safety and correctness.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by ΑΙhub.