Robotics is going through a new wave, driven by a novel form of artificial intelligence. At Tech For Industry Show 2026 (Paris Expo Porte de Versailles), a panel brought together NVIDIA, Exotec, Michelin and Dassault Systèmes to answer five key questions about physical AI. Starting point: of 550 warehouse and factory managers surveyed, 63% believe tomorrow's industry will be run or supported by AI, and 38% have already started this transformation.

For NVIDIA, physical AI "is when AI leaves your PCs and smartphones and enters the real world." For decades, robotics meant programming a robot for a precise task — but moving a bolt a few centimeters required recoding everything. Physical AI brings the ability to perceive, reason and adapt behavior to the environment. It rests on three pillars: embedded compute, simulation of real environments, and a foundation as open as possible (open source) to deploy with partners.

For Michelin, the stake is moving from "very high-performing but deterministic" automation to autonomy you want to control: a more flexible, learning factory, resilient to hazards — without falling into the science fiction of a 100% AI-run factory. Exotec confirms it from warehouses: physical AI meets the need for flexibility and the handling of degraded cases (the missing bolt in 1-2% of cases, which, over millions of operations, happens often). The best synthesis: moving from coding a robot to training it, while keeping deterministic where necessary (welding, assembly) and adaptive where the environment varies.

Despite the enthusiasm, maximum adoption isn't there. Several barriers compound.

Existing factories have old systems and poorly accessible data in heterogeneous formats — hence the answer of an open format like OpenUSD to consolidate them. To scale without running AGVs 24/7 at the risk of accidents (as with autonomous vehicles), you need simulation that's both fast and realistic enough (NVIDIA Omniverse). Finally, physical-AI experts with 20 years of experience don't exist: the field didn't exist 5 years ago. NVIDIA adds a cultural message: Europe, more risk-averse, must dare to explore — "to ride a bike, you have to ride a bike."

Michelin details the industrial and human barriers. Reliability first: a factory runs 24/7, requiring robust, maintainable solutions. Safety next, fundamental once an AI "takes physical form" near machines and people — "if you put a humanoid in a workshop, the first question is: where's the emergency stop button?". Human acceptance: a 1.80 m human-shaped robot in motion is intimidating, and having ten creates a scale effect to master. Skills finally: who will maintain and retrain these systems in a variable environment? Hybrid profiles (mechatronics, robotics, AI, data) that don't yet exist will be needed. Exotec recalls a mundane barrier: cost — paying more for what a classic controller already does 90% of the time is hard to sell; many will prefer a temp worker for the remaining 10%.

For Dassault Systèmes, automation "is absolutely necessary" — "taxing robots only slows automation and prevents keeping factories in France." But the factory mixes deterministic systems (conveyors, programmed robots) and autonomous ones: the real challenge is orchestrating all this, using autonomy "when it's useful," since it has a compute cost and time (controllers seek 3 to 5 milliseconds). Factory data exists only with whoever operates it: hence work with NVIDIA on combining virtual twins + synthetic data to train robots faster. "Simulation isn't cheating, it's preparing": if it doesn't work in simulation, it has no chance of working in reality.

Michelin observes "many proofs of concept" that must be exited: value isn't yet fully proven, it will depend on mass adoption and cost. PoCs address the functional (can the robot handle the load?) and adaptability ("the brain"), but the hardware — deemed "fragile" by mechanics — still needs work via a TRL increase. Exotec nuances: hardware is progressing fast (a French talent pool exists), and it's rather model maturity and the ability to relearn that's lacking (a human operator becomes effective in days, a humanoid requires weeks of training). Shared recommendation: don't wait for the perfect solution, tackle tasks where value and maturity allow it — "I don't need a humanoid to move a crate, a conveyor is enough."

Central debate. Dassault Systèmes recalls that the real breakthrough is perception (interpretation then action), useful for AGVs, AMRs and random-picking robots — which have no legs. NVIDIA embraces the humanoid as an R&D flagship (like autonomous vehicles: succeeding on the humanoid "drives" everything else), but defends a platform letting the most suitable form factors emerge for each use case. Concrete examples abound: Wandercraft deploys autonomous humanoids at Renault (they carry tires), Schaeffler in Germany is preparing some, and above all Pegatron (semiconductor factories in Asia) and Amazon (warehouse conveyors) show that scaling has already happened. A humanoid trained for weeks for a two-minute sequence becomes profitable once deployed across a fleet of 300-400 robots.

The shared conviction: no one succeeds alone. Michelin combines internal projects (developing its skills), partnerships with suppliers (understanding real maturity) and collaborations with research/academic institutes for very-low-TRL technologies. Dassault Systèmes recalls Europe has two major virtual-twin players (including Dassault Systèmes and Siemens) and works with NVIDIA, as well as Neura Robotics (whose "Neura Gyms" are physical training environments) and Agile Robots — the future being a combination of physical and virtual training (3DEXPERIENCE, Siemens platforms, NVIDIA building blocks). NVIDIA insists: its platform "is the ecosystem" (integrators, software vendors, startups, institutes, Accenture, and in France Mistral and others). The message: "let's all get on the bike," through co-development more than simple partnership.

Asked about one action to make Europe a leader, the speakers converge: NVIDIA bets on synthetic data and making "environments talk to each other"; Exotec on the hardware + software value chain; and Michelin, like Dassault Systèmes, on training — jobs are transforming (mechatronics + AI + simulation + integration), curricula aren't yet adapted, and there's no single "roboticist" profile. Michelin has even created an internal physical-AI academy to train and retrain its talent. Final message: "the moment is now," with industrial and national courage to assume.

What is physical AI? It's AI that "leaves the screens" to enter the real world via robotics: a robot's ability to perceive its environment, reason and adapt its behavior — instead of running a fixed program. You move from coding a robot to training it.

Why doesn't robotics scale yet? Because of compounding barriers: poorly accessible data, the need for realistic simulation, expert shortage, unproven 24/7 robustness, safety stakes, solution cost and the absence of hybrid profiles (mechatronics + AI + data).

Is simulation cheating? No: "simulation isn't cheating, it's preparing." If a solution doesn't work in simulation, it has no chance of working in reality. Combined with synthetic data and virtual twins, it lets you train robots at scale without risk.

Are humanoid robots ready for industry? They mainly serve as an R&D flagship. For immediate industry, it's rather legless form factors (AGV, AMR, picking robots) with perception that are deploying. Humanoids already exist (Wandercraft at Renault), but their generalization will take time.

This panel was recorded at Tech For Industry Show 2026, the B2B trade show for industrial transformation (Smart Manufacturing, AI, Supply Chain) at Paris Expo Porte de Versailles. Watch the full discussion above, and join us on May 19–20, 2027, Hall 7, for the next edition.