ABB Robotics, the robot-making arm of Switzerland-based ABB, is turning simulation into a lab for what-if experiments, from evening glare to dust and scratched parts.
Simulation once mainly checked that systems were ready to go live. Now it generates cheaper synthetic data.
“Synthetic datasets are not going to replace the real deployment, but if we reduce the time a robot needs to spend in the real world, that’s profitability,” said Boris Fiedler, head of operations and portfolio for software and digital at ABB Robotics. “It reduces risk and increases the payoff for the customer.”
Fiedler said no simulation perfectly copies the real world, at least not yet.
He said the industry is shifting toward using simulation to run experiments, such as testing how a vision system copes with evening sun. The resulting synthetic data, including pictures, videos and virtual robot forces, feeds the neural networks behind physical artificial intelligence (AI).
He said that when a deployment at a large fashion-industry customer goes wrong, ABB feeds the real data back into the original simulation. It then builds a richer set of synthetic variations around that single failure to train better AI.
“If you invest in simulation early and only think about the initial deployment, you’re kind of wrong because the system keeps working,” he said. “You will keep having a payoff every time there’s a new situation. There is a continuous payoff from having really good simulation on your systems.”
The technology now cuts across industries, from construction, where metal is bent and joined, to laboratory automation in life sciences. Lab tasks such as pipetting are simple, but one robot must handle many of them.
SoftBank Group agreed in October 2025 to buy ABB’s robotics business for $5.375 billion.
The unstructured frontier
Fiedler spoke at a fireside chat on digital twins and the gap between simulation and real life at the Physical AI and Robotics Summit, organized by Economist Enterprise in London on October 13. Laveena Iyer, a senior analyst at the Economist Intelligence Unit, moderated the session.
Fiedler cited kitting, where one worker handles a chain of tasks:
Opens the box
Removes the cushions
Takes out a part
Inspects it for defects
Puts it away
“Unstructured task is the frontier. Even though the environment is structured, the task itself is very dense, which is a massive challenge for automation,” he said.
The next challenge comes after deployment, when conditions keep changing, such as dust, a new stock-keeping unit (SKU) or defective parts. He said teleoperation, in which humans control robots remotely, is one way to close the gap, while ABB also uses its synthetic world to create more cases and more variety.
“I don’t think there’s one single answer, but the different technologies together will help to get it closer to generating the three nines in the end for the customer,” he said.
“Three nines” means 99.9% reliability. He said physical AI makes projects less predictable than traditional automation.
“We need to understand it more deeply. We need to work closer with our customers on certain problems, because problems and processes which we thought couldn’t be solved before can be solved today,” he said.
Consumer electronics poses a different test. Fiedler said the sector is now as heavily automated as automotive, yet some steps remain manual. ABB is working with Foxconn, the Taiwanese contract manufacturer behind many of the world’s smartphones, at one of its electronics manufacturing services (EMS) sites to raise yields further.
Tiny parts must meet extremely high quality standards, and product cycles are short. Once manufacturers finalize a new phone or tablet design, they usually have only about three months before launch to scale up the entire production line.
“If you have your phone in your hand, you want to have it perfect,” he said. “This is where simulation is so powerful, because you can do tons of this validation early in the process, so that the time you need to ramp up the production volumes is decreased significantly.”
Asia dominates robot installations
Global industrial robot installations rose 11% to a record 603,000 units in 2025, according to the International Federation of Robotics (IFR). China accounted for 59% of the total, and Asia for about three-quarters; installations fell 11% in the European Union (EU) and rose 12% in the US. Automotive installations in the EU fell 25% to 14,900 units.
“Why is Europe weak? Because automotive is weak, and there’s a huge transformation around the tier ones [suppliers], which struggle,” Fiedler said.
He said Europe remains strong in healthcare and other growing sectors that may eventually overtake automotive. Flexible physical AI systems are also growing fast, though it is early for this to show in the numbers. Random bin picking, in which robots pick randomly piled parts from a bin, has grown about 20% over the years.
“When you look into the industry, where growth is and what kind of application is equally important as the regions themselves, because that will tell you a lot about where investment is being directed,” he said.
He said scaling would remain a hurdle even if the gap between simulation and reality were largely closed.
“If we bring those [research systems] into pilots or production, it’s PhDs who are doing the work, and that’s not scalable,” he said. “The next big challenge is how to make this technology accessible to a robot engineer who is used to running around with a big screwdriver and bolting the metal to the ground.”
That means packaging autonomous systems so people who know the factory process, but aren't AI or software specialists, can deploy them.
His advice to companies planning a rollout is not to copy exactly how a human does a task, which he noted sounds counterintuitive, since autonomous systems are often expected to mimic humans. The process may not need as much redesign as in the past, because today’s systems adapt more.
“Very often that might not be the most efficient thing to do. Maybe if you change the process slightly, you can generate a much bigger return and have a much more effective production system,” he said.
His second piece of advice was to be open to agentic frameworks and “agentic engineering.”
“A simulation experimentation engine is the ideal tool for agentic engineering because the agents have something to work against, to test and to iterate quickly,” he said.
These tools aim to get companies into real-world testing faster.