- What does “physical AI” actually mean here?
- The term comes out of technology vendors’ marketing. NVIDIA’s glossary describes physical AI as letting autonomous systems perceive, understand, reason, and perform or orchestrate complex actions in the physical world. For operations we use a narrower reading: part of the movement decision comes from a learned model rather than from explicitly programmed logic. Everything else on this page follows from that one distinction.
- Is there a standard we can accept learned behaviour against?
- No. ISO/IEC TR 5469:2024 (“Artificial intelligence — Functional safety and AI systems”) describes properties, risk factors, methods and processes for using AI in safety-related functions, but as a Technical Report it is purely informative and carries no normative force. No binding acceptance or validation rule for learned robot behaviour exists to date. Operators close that gap themselves, through the operating envelope, the out-of-scope catalogue and scenario coverage in the contract.
- What is actually deployed today, and what is a demonstration?
- Epoch AI draws that line in its report “Where Autonomy Works: Evaluating Robot Capabilities in 2026”: navigation is in commercial use, for instance moving goods in warehouses, while most tasks requiring robots to handle, assemble or manipulate objects remain largely in the lab. Warehouse picking counts as the clearest case of commercially deployed manipulation in a controlled environment. Transfer of learned capabilities to new objects, environments and tasks is, according to Epoch AI, rarely demonstrated, yet it matters for most applications. As a market example, Amazon states that its Vulcan system handles around 75 percent of the item types stored in its fulfilment centres, in use at sites including Spokane, Washington and Hamburg.
- How do we tell whether a learned model is really driving?
- Three questions to the vendor. Which part of the system uses machine learning — only object recognition in the camera image, or the whole of trajectory planning? How is the movement computed — through fixed mathematical algorithms, or as end-to-end inference over network weights? And how is the behaviour verified — through detection accuracy plus a path check, or through statistical evaluation of hundreds of runs inside the operating envelope? The first case stays testable in the classical way. The second does not.
- When is deterministic programming still the right answer?
- In most cases. Where routes are clearly structured, processes are cycled and environmental conditions are controllable, deterministic software is more reliable, lower in latency and more predictable in operating cost. Learned components belong where unstructured load carriers or changing handover points demand flexibility. We specify the two alongside each other rather than against each other.
- Does werob build or train these models?
- No. werob is a manufacturer-independent systems integrator, not an OEM, not a manufacturer and not a model developer. We specify, select across more than 44 OEM partners, integrate into your systems and run the fleet in service. The models come from the manufacturers; our job is that acceptance, change state and the safety layer stay demonstrable for you.
- Is this the same as the Academic & Industrial R&D solution?
- No. That solution is about sourcing test hardware for a specific research project, for example a reconfigurable platform for autonomy testing. This page is about what changes in day-to-day operation when part of the behaviour of a productive fleet comes out of a learned model.
- How does this sit inside the operating model?
- The operating model itself is covered on the workforce page: how one job list ties together people, robots and software agents, and why the fallback path is the actual product. This page stays with the narrower question of what a learned model in the vehicle means for specification, acceptance and change control.
