TÜV SÜD says simulated robot safety checks could start production up to 30% sooner

Before an autonomous mobile robot can work in an aisle next to people, someone has to show it won't trap a worker. TÜV SÜD and Accenture say that at a medium-sized production plant, the rounds of coordinating, validating and signing off those safety concepts can take several weeks.

The two companies want most of that work finished before the robot is even delivered. At NVIDIA GTC Berlin, they presented the concept of a Physical AI Safety Suite. It uses digital twins, AI agents and NVIDIA simulation software to build the safety case on screen, then hands it to TÜV SÜD for independent review. Accenture's October 21 announcement cited a TÜV SÜD estimate that this digital approach could get production running up to 30% sooner. Technische Logistik reported the same estimate as speeding up commissioning by up to 30 percent.

The announcement describes the suite as partly digitizing three manual jobs: assessing facilities, assessing equipment and preparing certification packages. That cuts physical dependencies such as installed equipment and site visits, and lets companies fix safety and certification issues before they become costly rework in operation. The time saving appears to come from this decoupling. The evidence no longer has to wait for the hardware to arrive.

The inspector isn't going away. TÜV SÜD CEO Patrick Vollmer told Jasmin Jessen of Manufacturing Digital that he expects a "hybrid assurance model," with virtual and physical evidence backing each other up. Vollmer says digital twins and simulation "allow us to enter the physical inspection phase much better prepared." It's a measured position for an organisation that was founded in 1866 as a steam boiler inspection association and now has more than 30,000 employees at over 1,000 locations in about 50 countries.

In the announcement, Vollmer described what he sees changing. Physical AI makes robotic systems increasingly adaptive, he said, while also "paving the way for an independent, evidence-based digital certification process." For an operator, the real question is which parts of a safety case can credibly move into a model, and which parts still need a person with a clipboard on the warehouse floor.

Three layers, from floor plan to single robot

The suite splits the safety case into three connected layers. Accenture's announcement frames each one as a question an assessor would ask: where physical AI can be deployed safely, how it operates safely, and what it is allowed to do. TÜV SÜD provides the independent assessment at every layer and keeps responsibility for certification.

The first layer handles the site. It uses NVIDIA Omniverse libraries, the open Cosmos world model platform and the Isaac Sim robotics framework to check routes, clearances, mixed-use zones and areas where people and robots work together. Then it produces a structured evidence package for TÜV SÜD covering operating-zone requirements, hazard and risk reduction, robot integration norms and control-system safety. The announcement says this can run before robots are deployed or before layouts are reconfigured. That suggests the layer could also apply to existing sites that change their layouts, not only to new buildings.

The second layer simulates whole fleets to find operational limits and hazards ahead of deployment, with workers trapped by AMRs as the named example. It also keeps running after go-live. During operation, it relies on the NVIDIA Halos safety platform and the NVIDIA IGX Thor edge AI platform to spot hazardous situations outside any single robot's view. AI agents built on NVIDIA's Halos Outside-In Safety Blueprint use facility-mounted cameras to extend each robot's perception, so the fleet management system can react earlier.

That runtime piece changes what the product is. A tool that only prepares certification is finished once the inspector signs. Here, the second layer keeps contributing safety evidence for as long as the site runs, and Maitin's description of the suite says it "enables continuous monitoring of automated facilities."

The third layer covers the individual machine. Equipment that meets every product regulation still has to be tested under the conditions of its specific site: load, speed, floor surface and lighting. Using Isaac Sim and Cosmos, the suite generates and tests what the announcement calls a long tail of certification-relevant scenarios that real-world testing could never practically cover. One result the companies point to is that engineers can work out where safety sensors should go before a physical prototype exists.

Accenture's Sarat Maitin, who leads Supply Chain and Engineering for Central and Eastern Europe, explained the timing in the announcement. Companies too often test automation concepts against standards after equipment, layouts and workflows are already fixed, he said, and the suite moves that testing to a point where safety incidents and costly rework can still be avoided. Earlier testing and continuous monitoring, he added, both support faster deployment of the automation needed to raise manufacturing productivity, "particularly in Europe."

What still gets checked on the floor

In his Manufacturing Digital interview, Vollmer listed what on-site inspection still has to confirm. That covers how the sensors really perform, the tolerances of the hardware, the conditions of the installation, environmental influences, and the way people and machines actually interact. In other words, it covers everything a model might get slightly wrong once it meets the real building.

The handoff works in both directions. According to Manufacturing Digital, digital twins can also provide evidence covering rare or hazardous conditions, the kind that are hard to recreate on a real floor. Nobody wants to stage a worker getting pinned by an AMR just to prove the robot stops. So each method covers a gap the other can't. Simulation reaches the dangerous edge cases, and the floor inspection confirms that the sensors, tolerances and installation match what the model assumed.

Simulation moves much of the safety review into a digital model, so it can run before the robots go live on the floor. Physical testing still has the edge on one thing, and that's the real building. A warehouse whose floors, lighting or traffic patterns drift from the model will see that gap at inspection. The 30% figure is a TÜV SÜD estimate for a concept presented at NVIDIA GTC Berlin. That doesn't make the estimate wrong. It means the size of the gain will depend on how closely each digital twin matches its building.

The rulebook for simulation evidence

Vollmer was direct about the standards gap. Product standards such as ISO 3691-4 remain essential, he said, but they have to evolve as AI-based behaviour and new sensor technologies arrive on the floor. He also argued that fitting every requirement for physical AI, digital twins and simulation into one product standard won't be practical. Instead, complementary standards are needed to define how simulation evidence is created, documented and assessed.

He named three things those standards would have to cover: how accurate the digital twin is, whether its models and data can be traced, and whether it covers the relevant scenarios. Until such standards exist, the independent assessment is carried out by TÜV SÜD, which also remains responsible for certification.

TÜV SÜD's own pitch should be judged by the same three criteria. Vollmer argues that evidence produced in simulation lets teams prepare for compliance at a far earlier stage of engineering, "without compromising the rigor safety standards require." A buyer or an assessor could use accuracy, traceability and scenario coverage to check whether that rigour actually holds for a given twin.

That's our editorial suggestion, not a reported requirement. It follows directly from the criteria TÜV SÜD's CEO laid out.

Warehouses go first, within 12 months

Maitin's case for moving now starts with a ceiling. In Manufacturing Digital's interview, he cited Accenture research finding that even the most advanced factories run at roughly 50% to 70% automation. The rest is work that fixed, rule-based robots can't handle, and that's the work physical AI is aimed at. Manufacturing Digital also cited the IFR's World Robotics 2025 report and Deloitte's "Physical AI: The moment of acceleration" report, which show demand building in logistics while broad adoption is still at an early stage.

Accenture's announcement defines physical AI as a mix of technologies, including digital twins, simulations, sensors, robots and AI models, that lets machines sense, decide, act and improve constantly under changing conditions. That adaptability is what lets them take on the remaining 30% to 50% of tasks. It also helps explain Vollmer's push for new standards: a robot that keeps changing its behaviour is harder to sign off once and forget.

A safety sign-off that takes several weeks per site is manageable for a pilot. As companies scale from proof of concept to large-scale rollout, Maitin says, production will come online sooner if safety assessments are part of planning instead of being left until the end. Over the next 12 months, he expects the first large-scale deployments to arrive, with physical AI scaling first in warehouses and logistics and then in manufacturing.

Accenture has also set up a Center of Excellence that builds and deploys physical AI systems for warehouses and production sites. Maitin gives operators blunt advice: companies that want a competitive edge need to begin bringing in physical AI today, rather than holding off until the technology evolves further.

For warehouse operators planning AMR rollouts in that 12-month window, this moves the safety case to the start of the project. If the assessment begins with a simulated layout, the floor plan, routes and shared zones would have to be modelled before the hardware order rather than after installation.