At Automate 2026, Bucket Robotics demonstrated more than another AI-powered inspection system. It offered a glimpse into a future where machine vision systems can be trained before a production line even starts running, dramatically reducing deployment times and opening the door to applications that were previously uneconomical.
Walk around any manufacturing exhibition today and it doesn’t take long before someone mentions artificial intelligence. The phrase has become almost unavoidable. Every camera manufacturer, software developer and systems integrator is keen to explain how AI is transforming inspection, automation and quality control. Yet behind the marketing, one problem continues to slow adoption.
Data.
Not algorithms. Not processing power. Not even cameras.
For manufacturers looking to deploy AI-powered inspection, the biggest hurdle is often gathering enough representative images to train a reliable model. Every product variation needs examples. Every defect needs documenting. Lighting conditions have to be accounted for, production tolerances understood and thousands of images painstakingly labelled before an inspection system is ready to enter service. For manufacturers producing a limited range of products, that’s challenging enough. For companies manufacturing hundreds—or even thousands—of different components, it can become almost impossible.
It is a problem that has existed almost since AI first entered industrial vision. Bucket Robotics believes it has found a solution.
Speaking to MV Pro at Automate 2026, CEO Matt explained how the company is using synthetic data to remove what he describes as the “cold-start problem” in machine vision. Rather than beginning with photographs collected from the factory floor, Bucket Robotics starts with something manufacturers already possess: the CAD model.
Using those engineering files, the software creates a digital representation of the component before generating synthetic training images that include not only perfect parts, but simulated defects, changing viewpoints and varying environmental conditions. Instead of waiting weeks or months to collect sufficient production images, manufacturers can begin training inspection models immediately.
“We’re solving the cold-start problem for quality inspection.”
That simple statement captures what could prove to be one of the most significant shifts currently taking place in industrial vision. For years, synthetic data has been discussed as an interesting research topic. Companies have demonstrated impressive examples of simulated environments and digitally generated defects, but many manufacturers remained cautious. Simulation sounded promising, yet production engineers are naturally sceptical of anything that hasn’t survived the realities of a factory floor.
This year, however, the conversation appears to have changed. Rather than asking whether synthetic data works, visitors to Automate were increasingly asking how quickly it could be deployed.
That distinction matters.
It suggests the industry is moving beyond experimentation and towards practical implementation. Bucket Robotics has noticed the change first-hand. The company originally focused on quality inspection, helping manufacturers identify defects using AI trained on synthetic data. Increasingly, though, customers are asking for something much broader. Once manufacturers understand that new inspection models can be created rapidly, they begin looking beyond a single end-of-line quality station. They start asking whether the same approach could verify assemblies halfway through production. Whether it could count components before a machine closes. Whether it could confirm that operators have completed every stage of a manual assembly process.
In other words, the conversation shifts from quality control to process validation.
That represents a much larger opportunity.
Traditional machine vision has often been reserved for applications where the return on investment could justify lengthy integration projects. A single inspection station at the end of a production line made sense because it protected the value of the finished product. Inline inspection has always been desirable, but rarely economical. Synthetic data has the potential to change that equation. If new inspection models can be generated in hours rather than months, manufacturers suddenly gain the flexibility to deploy vision systems throughout production rather than only at the final quality gate. It is perhaps one of the least obvious consequences of simulation, yet potentially one of the most significant.
Another area where Bucket Robotics believes synthetic data offers a major advantage is environmental robustness.
Lighting has always been one of machine vision’s greatest challenges. A system that performs perfectly during commissioning can quickly become unreliable if factory lighting changes, shadows fall across a component or reflective surfaces behave differently from one production line to another. Rather than treating those variations as problems to solve after installation, Bucket Robotics introduces them during training.
Thousands of simulated lighting conditions can be generated automatically, exposing the AI model to situations that may only occur occasionally in real production. Instead of simply recognising defects, the software also learns what normal variation looks like.
Can it distinguish between a shadow and a burn mark? Will it recognise the same component under different lighting conditions?
These are exactly the types of questions that traditionally required lengthy testing on the factory floor. Simulation allows many of them to be answered before deployment.
“By spending so much time in simulation, you can train the system to understand whether it’s looking at a shadow or a genuine defect.”
Interestingly, Bucket Robotics chose not to showcase this technology with an elaborate exhibition stand. Matt jokingly described the company’s demonstration as “the least sexy vision demo at Automate.”
He may have had a point.
Visitors expecting dramatic robot choreography instead found cable chains, serial numbers and assembly verification. Components were deliberately presented under realistic conditions rather than perfectly controlled laboratory lighting. Yet that understated approach arguably reinforced the company’s message. Factories are rarely tidy. Operators move components by hand. Shadows appear unexpectedly. Camera angles change. Inspection systems need to function in that environment rather than inside carefully staged demonstrations. Perhaps the most striking claim made during the interview concerned deployment speed.
Manufacturers frequently report AI inspection projects taking several months to reach production. Image capture, annotation, testing and optimisation often become the longest phases of the project. Bucket Robotics argues that much of this work can be completed before a single production image is collected. If customers can provide CAD data and define what constitutes an acceptable component, the company says it can begin generating inspection models almost immediately. Real-world production images remain valuable, but they become a method of refinement rather than the foundation of the system.
For manufacturers producing thousands of different components, that distinction could fundamentally alter the economics of automation. Collecting representative defect images across an entire product catalogue is often unrealistic. Some defects may occur only rarely, making it impossible to gather sufficient examples without waiting months—or even years.
Synthetic data removes that dependency. Instead of hoping defective parts eventually appear, manufacturers can simply generate them digitally. The implications extend well beyond faster deployment. They also change how inspection projects scale.
Many manufacturers begin their AI journey cautiously, installing a single inspection station before expanding across the factory. If creating new inspection models becomes largely a software exercise rather than a lengthy data collection project, expansion becomes considerably easier.
During the interview, Matt described customers moving from individual component inspection towards complete assembly verification and even maintenance applications. The same simulation-driven workflow can support entirely different inspection tasks without requiring months of fresh data collection every time.
Perhaps the best illustration came from a simple experiment. A colleague selected three components from the McMaster-Carr catalogue. The parts were ordered late in the day. By the following morning, Bucket Robotics had already built inspection demonstrations covering counting, defect detection and assembly verification using synthetic data generated before the physical components even arrived.
It is an example that neatly illustrates how quickly this approach can move from concept to deployment. For the wider machine vision industry, that may be the real story emerging from Automate 2026. Synthetic data is no longer simply about creating artificial defects. It is becoming a new way of thinking about inspection altogether. Instead of beginning with the camera, manufacturers begin with the digital model. Instead of collecting thousands of production images before deployment, they generate them automatically. Instead of waiting months to validate an application, they begin testing almost immediately.
None of this suggests that real production data will disappear. Factory validation will always remain essential, and no simulation can perfectly recreate every variable found in a live manufacturing environment.
What appears to be changing is the balance.
Real-world data is becoming the final stage of development rather than the first. That subtle shift has the potential to reduce costs, shorten deployment times and make AI inspection practical for applications that would previously have been difficult to justify. For Bucket Robotics, the challenge now is turning that promise into widespread adoption. For the rest of the industry, the message from Automate was equally clear.
Synthetic data has moved well beyond the research lab. It is beginning to reshape how manufacturers think about machine vision itself, and that could prove to be one of the defining trends of the next generation of industrial automation.
