I am a Physical AI consultant and project manager specialising in autonomous systems, robotics and embodied intelligence. After a career spanning large-scale industrial projects across Energy (Oil & Gas), Finance (CIB/Trading) and IT, I made a deliberate transition toward Physical AI, the frontier where artificial intelligence meets the physical world. I lead the LOGIE-AI consortium (France 2030), advise large industrial groups on GenAI and AI programme delivery, and hold three MIT certificates in applied data science, machine learning and physical AI. I am completing the OpenUSD certification.
Research & Publications
Peer-reviewed papers, white papers and technical reports on Physical AI and autonomous systems
Robotics
Physical AI Frontiers — The Pause That Sounds Like Thinking: Conversational Timing in Social Robots
July 2026
Why human turn transitions average 208 milliseconds, why one laboratory study found users preferred about 700 milliseconds from a robot, and how endpointing, not model speed, dominates the delay people actually experience in conversational robotics.
Physical AI Frontiers — The Last Hundred Microns: Precision Assembly When Clearance Is Smaller Than Workcell Uncertainty
July 2026
When the combined uncertainty of robot, tooling, fixturing and parts exceeds assembly clearance, precision has to be recovered through the interaction loop: the geometry of jamming, admittance control, compliant end-effectors, and simulation-trained insertion policies that now transfer zero-shot at industrial tolerances.
Physical AI Frontiers — The Reality Gap: Simulation-trained policies, synthetic data, and what actually transfers from the simulator to the factory floor
July 2026
How simulation moved from test bench to training data factory, why the gap between simulated and real behaviour is now the central engineering risk in robotics, and what managers need to know about sim-to-real transfer, learned dynamics and scaling laws for dexterity.
Physical AI Frontiers — The Last Inch: Kinesthetic Sensing, Dense Clutter, and the Quantity a Camera Cannot Measure
July 2026
A field analysis of Amazon's contact-rich stow robotics deployment: how kinesthetic feedback corrects vision-only estimates, why compressibility is the variable no camera can measure, and what it means for industrial project managers buying embodied AI.
Telemetry analytics, digital twins and bounded autonomy in the risk architecture of a satellite fleet: how ML anomaly detection, operational twins and tiered machine authority are moving spacecraft risk management from insurance transfer toward engineering resilience.
Physical AI Frontiers — One Scene, Many Solvers: OpenUSD as a Foundation for Physical AI Simulation Pipelines
July 2026
Why a shared scene description layer is becoming the connective tissue of robotics simulation, synthetic data and sim-to-real transfer, and why standardised description does not mean standardised physics: what composition reliably buys you, where solver semantics diverge, and how to adopt OpenUSD without believing the marketing.
Digital Twin Series — The Number You Cannot Audit: Predictive maintenance and digital twins in power generation
August 2026
Four classes of predictive-maintenance KPI, ranked by how well each survives scrutiny, and why avoided cost is model-dependent rather than measured. A gas-turbine compressor fouling case shows what can be proven from instrumentation most plants already have, and what cannot.
Digital Twin Series — The Cables Were Never the Problem: Three electrical lots on a waste-to-energy build, and why configuration governed the schedule
August 2026
On a 1.8 billion euro waste-to-energy programme, the physical work landed broadly on plan while plant-specific functional engineering drove the critical path. Why a plant's intended behaviour gets specified later and less rigorously than its geometry, and where virtual commissioning moves that discovery upstream.