Physical AI must become industrial before it can scale
Robots are beginning to move beyond fixed paths and tightly controlled tasks, but manufacturing will judge the next generation of automation by a familiar standard: whether it works reliably enough to justify its place in production. Physical AI can give robots more awareness of their surroundings and allow them to respond to variation that conventional systems struggle with. That capability only becomes valuable when it continues to perform through changing factory conditions and delivers the return expected from industrial equipment.
Craig McDonnell, Business Line Managing Director, Industries at ABB Robotics, says the challenge now is to turn years of experimentation into something manufacturers can deploy with the same confidence they expect from conventional automation. “We need industrial-hardened physical AI. The automation industry has been conditioned to expect a certain degree of reliability and consistency, and the business cases used to fund automation are built around that. There are many examples of physical AI today that do not yet live up to those expectations. Unless you can deploy these systems reliably and repeatedly, it is very difficult to make the economics work for the manufacturer or the end user.”
ABB expects physical AI to widen the range of work that robots can perform as sensing improves and motion becomes more autonomous. McDonnell also sees a more immediate opportunity in revisiting existing applications where small amounts of variability still cause lost productivity. The path to a larger robotics market may therefore begin with making familiar automation considerably more tolerant of the real factory rather than waiting for completely general-purpose machines.
Closing the sim-to-real gap
Simulation is central to that effort because manufacturers cannot collect every possible production condition before commissioning a system. The problem has been the gap between what happens in the digital environment and what the robot encounters once the application is installed. Small errors in the simulation can become much more significant when they interact with the tolerances of the physical robot and its vision system.
ABB has been working with NVIDIA to enrich synthetic data and create more detailed digital representations so that more of that uncertainty can be addressed before the cell reaches production. “We have had simulated environments that were around 80 percent accurate, and that is not enough to reliably build and pre-train an automation system,” McDonnell says. “Working with NVIDIA, we have managed in structured applications such as electronics and material handling to get up to around 99 percent correlation between the synthetically generated path and the real robot path. Once you reach that level, you can vary material tolerances and lighting conditions, then pre-train the vision system and make sure the application continues to operate when those conditions change in the factory.”
RobotStudio HyperReality extends ABB’s RobotStudio environment with NVIDIA Omniverse technologies so that engineers can model more of what surrounds the robot rather than concentrating only on its motion. Digital representations of the part and vision system can be exposed to changing environmental conditions before the physical cell is available, allowing more of the engineering work to take place earlier.
A manufacturer could begin developing an automation application before the final physical part arrives, test synthetic versions against different conditions and identify weaknesses before factory acceptance. McDonnell says this opens the door to more concurrent engineering, reducing the period traditionally spent tuning a cell after the equipment and product finally come together.
Digital-first validation is not intended to replace real-world testing. Its value is in moving more risk identification upstream, before the cell reaches the factory floor. A digital POC can establish whether the task, data and model are viable, but production readiness still has to be proven through a pilot cell and then through wider deployment, with real-world feedback continuing to refine the digital environment and the model.
Reliability must persist after commissioning
Factory adoption also depends on defining the limits of autonomy. McDonnell is clear that more intelligent perception does not transfer responsibility for safety to the AI model. “The day when we move away from a rules-based safety environment is still some way off,” he says. “Everything we are talking about needs to operate within a very clear functional-safety framework. We do not see that changing in the immediate future because the safety capability is not there yet. In an industrial-hardened environment, you still need the conventional risk assessment and safety rules around the application.”
Commissioning does not freeze an AI system in a known state either. As users introduce new data, the application itself can change, which means performance must be managed throughout its working life.
McDonnell gives the example of parcel handling, where users may continually add images, so the system learns to recognize new packages. “If the end user accidentally uploads bad images, the performance of the model can deteriorate,” he continues. “You might start with 99 percent reliability and then lose four or five percentage points simply because poor data has entered the model. It is not static, so you must manage the AI instance throughout its life.”
ABB’s AI Robot Trainer is intended to make that management accessible to the people running the application. Users can identify model deterioration and decide whether to remove problematic data or return to an earlier model version without requiring an AI specialist to be permanently present. McDonnell argues that the operator needs expertise in the manufacturing process rather than expertise in model development if physical AI is to become manageable at scale.
The immediate automation environment must be considered in the same way. Conveyor states and PLC information can be brought into simulation so that abnormal situations are tested before they happen in production. A dropped component, for example, can become a scenario the robot has already been trained to handle instead of an event that leaves material accumulating until somebody intervenes.
Value will arrive before general-purpose robotics
McDonnell expects value to spread outward from known automation problems rather than arrive first through completely general-purpose machines. “We are seeing value today in traditional structured environments by handling part and environmental variability much more effectively,” he says. “Then you move into repeatable processes that are more difficult and require much deeper vision integration. Beyond that are the highly unstructured applications. Those are getting very close, but they are not quite there yet.”
McDonnell points to machines at large electronics manufacturing services companies where reliability has risen from around 40 to 50 percent into the high 90s as HyperReality has been used to manage variability more effectively. Vision-intensive applications such as item picking and parcel handling are also progressing quickly, while more unstructured tasks, including mixed e-commerce returns and cable handling, remain closer to the development frontier.
Physical AI therefore does not need to culminate in a general-purpose humanoid before manufacturers see meaningful returns. ABB expects industrial robots to become more mobile and versatile while retaining forms designed around the work they need to perform. The strongest business cases are likely to remain applications where additional autonomy solves a defined production problem, and its performance can be measured against an existing process.
The same push toward flexibility could change who can improve an automated process once it is running. Static systems can leave shop-floor workers dependent on specialists whenever something needs to change, even when the people operating the line can see what would make it work better.
“If you are on the shop floor and somebody else has defined the process for you, you can see that it is not working and you cannot do anything about it, that is a terrible experience,” McDonnell adds. “If you can see how to improve it and interact directly with the machine, you become much more engaged. That is where AI can make automation more enabling for the people who run the process.”
Manufacturing will ultimately judge physical AI less by how impressive the intelligence appears than by whether the resulting system behaves like industrial equipment. Simulation can reduce commissioning risk, and better perception can make automation less brittle. Neither removes the need to prove performance under real production conditions, which is where physical AI will have to earn its place.
