Vision & Verification
AI vision raises the bar in EV battery assembly
Atlas Copco's VisionTools platform brings AI-enhanced machine vision to EV battery production, combining neural networks with traditional inspection logic to achieve unprecedented quality assurance at speed and scale.
The temptation, when confronted with a powerful new technology, is to overstate its scope. Machine vision in automotive production is no exception. When Atlas Copco's VisionTools team talks about artificial intelligence in factory inspection, the first thing they want to make clear is what AI does not do.
Alex Hunt, Vision Tools Engineer at Atlas Copco, is precise on the point. "AI is just one cog in a much larger machine," he says, "used for very specific tasks where a neural network is genuinely needed. Most assemblies contain many features that traditional vision tools can already inspect." The practical implications of this qualification are important: the system's architecture is deliberately hybrid. Neural networks - despite the vehicle production industry's general inclination to deploy across the board - are deployed selectively, in situations where rule-based approaches run out of road, and the two paradigms work in concert rather than competition.
AI is just one cog in a much larger machine, used for very specific tasks where a neural network is genuinely needed. Most assemblies contain many features that traditional vision tools can already inspect
Further; this is not false modesty. It reflects a considered engineering philosophy, and one that has direct consequences for how the VisionTools platform is specified and utilised. Before arriving at any discussion of capability, Hunt makes two important preliminary observations. First, traditional vision systems are not going anywhere. And second, the widening skills gap, now prevalent throughout automotive manufacturing, means that many production facilities simply do not have a dedicated vision engineer on hand to manage and troubleshoot systems as they evolve.
The vision engineer of 2026 is an unusually stratified figure. "A modern vision engineer now needs to understand a huge range of disciplines, including cameras, optics, lighting, electronics, networking, CAD, mechanical design, cabling, programming, logic, safety, cycle-time calculations, and more," Hunt explains. "It's a broad and demanding role." That breadth has shaped how VisionTools has designed both its technology and its customer support model, aiming to reduce the dependence on rare expertise at every possible point.
The optical challenge of EV battery inspection
However, despite the prior qualification that "most assemblies contain many features that traditional vision tools can already inspect", EV battery assembly is a growing exception, as it presents a set of inspection conditions that would defeat most standard imaging approaches.
Dark housing components, black sealants deposited on dark substrates, and transparent adhesives that are virtually invisible to conventional 2D cameras combine to create an environment in which contrast - the subtle and foundational requirement of machine vision - is perpetually scarce.
Hunt describes the toolkit his team employs to navigate these conditions. "Several techniques help separate visually similar materials," he explains, "such as low-angle lighting to highlight black sealant on dark surfaces, fluorescent lighting to reveal transparent adhesives, and spectral imaging to differentiate materials that appear identical to the human eye. In addition, multi-angle shadow casting is used to create composite images where single shots fail. We also offer in-house feasibility studies to determine the best approach for each customer's product."
The referenced feasibility study is not a mere formality. The combination of optics, lighting geometry, and inspection logic must be engineered together, and experience shapes the outcome as much as equipment selection. Hunt is blunt about the implications for customers approaching the technology with unrealistic expectations: "Sometimes the simplest truth is this: it may be easier to fix the manufacturing process than to buy a vision system to catch the failures." It is a proposition that takes a certain confidence to deliver, and it says something important about how Atlas Copco positions itself in the market.
ROBOcam and the engineering of motion
In automotive manufacturing, vision systems for quality control have taken on variegated forms, as both OEMs and tier suppliers are racing to develop multifarious means to achieve the same outcome: enhanced quality inspection. So looking at Atlas Copco's solutions, it strikes you that mounting a vision system on a robot arm, surely must introduce its own distinct category of engineering problems. Firstly, the camera is in constant motion, which means the field of view shifts continuously, and to boot, the resulting vibrations pose a persistent threat to image fidelity. Surely? AMS probed the seeming problem. The answer: Atlas Copco's ROBOcam (which was designed expressly for robot-mounted inspection in automated assembly ) was built around these challenges realities from the outset.
Hunt's enthusiasm for the product is evident. "ROBOcam is one of our favourite products," he says. And the engineering rationale becomes quickly apparent. Global shutter sensors capture the entire image simultaneously, eliminating the rolling distortion that afflicts cameras relying on sequential line capture when a subject is in motion.
Lighting intensity and shutter speed are, we are informed, calculated against expected motion blur. Furthermore, automatic zoom lenses adjust focus across varying fields of view and working distances without requiring manual recalibration. In other words, both the problems and the solutions of machine vision are approach from multiple angles at the same time.
Safety, however, was the overriding design consideration. As more robots, both collaborative and traditional industrial arms, are paired with vision systems, the proximity of operators introduces risks that must be engineered out - rather than managed around.
"The Vision Tools team designed curved, impact-reducing camera and lighting enclosures for robot end effectors," Hunt notes. It is no secret that collaborative robots require formal safety calculations, and the physical design of the ROBOcam enclosure reflects that obligation directly.
Asked what the single biggest engineering challenge was, Hunt's answer is characteristically precise: "The biggest challenge is usually optics selection, and we've engineered around that too." The remark speaks to a design process that took nothing for granted.
Training data: a question of context
Among the most persistent questions surrounding AI-based inspection is the volume of annotated training data required before a system can reliably distinguish a genuine defect from acceptable process variation. Hunt's answer is realistic in its perspective: the question, as commonly posed, is too broad to be useful. "In short, it doesn't underpin the whole verification process," he says. "Perhaps a more appropriate question is: 'what exact feature is the AI looking at?' Without that context, it's like asking, 'How long is a piece of string?'"
And the distinction matters in practice. A white sticker on a dark background is a task that a non-AI system handles without difficulty. A tiny white ten-millimetre sticker in a large, dark field of view, is primarily an optics and resolution challenge. While for foreign object detection, the variables multiply rapidly: a large part on a plain background requires minimal training, while the same part on a complex background requires more - and a small component on a complex background requires significantly more still.
If a small bolt lands perfectly on top of another bolt, the system may pass it. In such cases, multiple cameras or height-sensing technologies can help remove ambiguity
Hunt also draws attention to the edge cases that no training regime can entirely eliminate. "If a small bolt lands perfectly on top of another bolt, the system may pass it. In such cases, multiple cameras or height-sensing technologies can help remove ambiguity." It is a candid acknowledgement that the limits of any single detection modality are real, and that system design must account for them.
Deployed performance, however, is consistently strong. In well-engineered vision applications, Hunt reports, detection rates above 99% are common, typically delivered with greater consistency than manual inspection can achieve. "The key advantage over manual checks is repeatability," he says. "Human inspection can be very capable, but it is naturally affected by fatigue, shift changes, and subjective judgement." The traceability dimension reinforces this advantage further. Images and inspection data stored at every station allow manufacturers to analyse trends over time, investigate recurring issues, and continuously improve the process.
Temperature as a quality signal
For EV battery assemblies, an undetected defect carries consequences that extend well beyond a failed quality audit. Thermal runaway events can produce catastrophic outcomes, and the conditions that enable them are frequently laid down during assembly rather than in operation. Atlas Copco addresses this through the Advanced Verification offering built around the V60 software platform and integrated thermal imaging.
And the approach is methodical. "Thermal images are displayed using colour-scaled temperature maps," Hunt explains. "Users set upper and lower limits to trigger alarms. These thresholds are typically defined during process validation and adjusted to the specific battery chemistry, materials, and assembly process." The V60 platform connects to the customer's manufacturing execution system, providing centralised management of all cameras, lights, and accessories from a single interface.
The defect modes that thermal imaging is designed to catch are numerous: inconsistent adhesive or sealant thickness, inadequate cooling in cell or module housings, misalignment between conductive materials, and fastening irregularities that generate localised heat. Even temperature differentials too small for a human inspector to detect can compromise battery safety and performance if left unaddressed. The effect of integrating thermal imaging inline is to transform it from a retrospective check, applied after a battery pack is complete, into an active preventative control that operates continuously throughout assembly.
Adapting to new variants without downtime
Battery cell formats are evolving rapidly. Cylindrical, prismatic, and pouch geometries are all in active production, new chemistries are entering the supply chain on an accelerating timeline, and manufacturers face mounting pressure to ramp new variants quickly without disrupting existing lines. The ability to reconfigure an inspection programme fast is, in this context, a commercial capability as much as a technical one.
Hunt describes a process that sidesteps the conventional training cycle almost entirely. "As long as the user has a CAD model, they can generate synthetic images, build and validate programs offline, and simply load them when ready. There is no need to run the new variant online, capture images, or re-test everything from scratch." The approach has been validated in a live customer environment and is being developed further as part of a formal product initiative conducted in collaboration with a university partner.
At gigafactory scale, where model changeover downtime can cost millions, the practical value of this capability is substantial. It also signals a broader shift in how inspection programmes are developed and maintained, away from the traditional model of iterative trial-and-error on the production line and towards a front-loaded engineering process grounded in digital twins.
For VisionTools, our core focus is assembly verification. The goal is not to push multiple products into the same space, but to create a coherent architecture around customer needs
From detection to intervention
Further to this, detection alone is insufficient if the production line cannot act on what is found. The intervention loop that follows a defect detection in a VisionTools installation is shaped by the customer's process, but the platform is designed to support a range of response architectures.
Simple pass/fail logic travels as a digital output to a programmable logic controller. Richer data flows over network protocols. Results can include make, model, serial number, colour, and variant, and can be written to a part's unique data file, a practice that is now standard in high-integrity automotive manufacturing.
For correctable defects, the response can be immediate and localised. "If a defect can be corrected in place, such as removing a detected foreign object, the operator can be alerted via illuminated beacon or display, press a button, and re-inspect without removing the part from the line," explains Hunt. "Some customers use rework loops, where parts leave the main line until fixed and re-inspected."
The system's modularity is central to this flexibility. "Our AI module is part of a larger modular system, not a standalone island," Hunt notes. "It works alongside traditional tools and communication protocols to create a more powerful combined solution." The architecture ensures that AI-driven decisions do not operate in isolation but feed directly into the plant's existing quality and production management infrastructure.
Acquired vision: the corporate acquisition that changed the picture
Atlas Copco acquired VisionTools GmbH in late 2024. Adnan Eker, Global Business Line and Product Manager for Vision Tools, describes what the acquisition unlocked in terms of capability amplification rather than simple technology addition.
AI is better suited to offline quality control of cells, as well as the inspection of battery pack assemblies and subcomponents
"VisionTools brought strong expertise in assembly verification and AI-supported vision applications, while Atlas Copco adds a global customer base, a strong service footprint, and a broader smart manufacturing ecosystem," he explains. For EV manufacturers, the practical consequence is access to integrated quality concepts that connect machine vision directly with the wider assembly process, supporting traceability, risk reduction, and global scalability in a way that neither party could have delivered independently.
The portfolio governance question is one that Eker addresses without evasion. Within Atlas Copco, each product company operates within a defined focus area, and a common strategic framework is applied to prevent overlap between adjacent businesses. "For VisionTools, our core focus is assembly verification," he says.
"The goal is not to push multiple products into the same space, but to create a coherent architecture around customer needs. That requires internal alignment, clear ownership, and close collaboration across the different businesses." For an EV manufacturer engaging Atlas Copco as a strategic partner, the implication is clear: the inspection solution offered will be designed around the application rather than assembled from whichever product happened to be developed most recently.
Why battery assembly was the natural starting point
The decision to focus the VisionTools AI capability on EV battery assembly ahead of other automotive quality applications was driven by a specific set of converging market pressures that Eker traces with precision.
"The tipping point," he explains, "was essentially the increasing complexity and speed of battery production. EV battery assembly combines high throughput, tight safety requirements, and many applications where traditional rule-based inspection starts to reach its limits. At the same time, manufacturers are dealing with frequent product changes, new variants, and the pressure to ramp up faster. AI offers a way to make inspection systems more flexible and more scalable in those environments. It helps reduce the dependency on deep vision expertise for every adjustment and allows faster adaptation to real production conditions. That made battery assembly a very relevant starting point."
Hunt adds an important qualification on the near-term limits of inline AI at cell level. High-speed cell inspection, he notes, makes widespread inline AI adoption unlikely until cycle times improve materially. "AI is better suited to offline quality control of cells, as well as the inspection of battery pack assemblies and subcomponents." The honest delineation of where the technology does and does not yet perform is consistent with the team's broader approach: capability claims are kept close to demonstrable reality.
Customers also need stable processes, good data, and solutions that are easy to maintain and use on the shopfloor. AI can close an important part of the gap, but true zero-defect production requires the right combination of vision technology, process integration, and operational discipline
The zero-defect horizon: ambition and its constraints
Eker is measured in his assessment of where the industry currently stands in relation to the aspiration of zero-defect manufacturing. "Most customers are clearly moving in that direction, but they are at different levels of maturity," he says. "In many operations, especially where product variety is increasing and manual work is still involved, there is still a gap between quality ambition and what can be controlled consistently in daily production."
The obstacles, he is careful to note, are as much operational, as they are technical. Inspection technology must be robust, adaptable, and reliable, but that is necessary rather than sufficient. "Customers also need stable processes, good data, and solutions that are easy to maintain and use on the shopfloor. AI can close an important part of the gap, but true zero-defect production requires the right combination of vision technology, process integration, and operational discipline."
Looking three to five years ahead, Eker expects the most significant development to be a transition from detection towards active process integration. Inspection systems will increasingly function as nodes in a closed quality loop, feeding data upstream to support earlier intervention and defect prevention rather than simply flagging failures after they occur. "That is exactly where we want VisionTools to be positioned," he says. "Not only as a detection system, but as part of a smarter, more integrated quality concept."
Safety, liability, and the case for traceability
Battery safety is ultimately a public issue, not merely an industrial one. Thermal runaway events attract regulatory scrutiny, damage brand reputations, and in the worst cases, endanger lives. The documentation and compliance requirements that follow are correspondingly demanding, and the standard of evidence expected is rising.
"Battery safety is a very serious topic, and manufacturers need both reliable detection and clear documentation," says Eker. "AI-based verification can strengthen quality assurance by identifying anomalies earlier and more consistently, especially in complex production environments."
The integration of thermal imaging into the wider inspection concept adds another layer: continuous monitoring of heat development during assembly, with the capacity to detect abnormal conditions and trigger responses before a compromised unit enters the next stage of production. The traceable inspection records generated by the system are not a secondary benefit. They are, for many customers, the primary requirement.
A strategic partner, not an equipment supplier
Eker is explicit about the distinction Atlas Copco draws between supplying technology and acting as a strategic partner. "Being a strategic partner means starting from the customer's production challenge, not from a product catalogue," he says. "We work to understand their process, the failure modes, the quality risks, and the operational targets, and then define the right solution around that."
In practice, this means taking responsibility not only for the inspection equipment but for its integration into the customer's manufacturing environment, and for sustaining the quality concept as production programmes evolve and plants scale globally. For greenfield gigafactory builds in regions where local vision expertise may be scarce, this approach is not a differentiator but a precondition. "For greenfield deployments, this is critical, because customers need a robust technical solution and confidence that support will be available as production ramps up."
The concluding thought belongs to Eker, and it captures both the ambition and the sobriety that characterise the VisionTools proposition in its current form. "AI in machine vision is not about replacing industrial know-how," he says. "It's about making inspection more adaptable, more scalable, and more practical for real production environments. What is especially exciting today is that we are only at the beginning. As manufacturers continue to push for higher quality, greater flexibility, and faster ramp-ups, the role of vision systems will only become more important."
For a technology still in its early phase of adoption, that is both an accurate assessment and, given the pace at which EV battery production is scaling, a significant industrial opportunity.
