B&R, ABB's machine and factory automation division, has moved deep-learning vision processing off the inspection PC and into the machine controller. The company's new AI Smart Camera portfolio synchronizes cameras with motion axes at sub-microsecond precision, and a just-in-time compiler running on a quad-core processor cuts measurement-task processing time by up to 75 percent. ABB describes the integration — vision data feeding real-time control loops — as a first for industrial automation.

From parallel systems to one loop

Conventional architectures keep vision and control in separate worlds. Bolting a camera onto a machine already managing safety functions, motion, robotics and CNC consumes engineering resources, and inspection results reach downstream decisions rather than machine behavior. Machines "fly blind," in B&R's phrasing. The new portfolio replaces standalone vision sensors and PC-based processing with edge-driven applications that run AI and rules-based inspection without interrupting production.

The Smart Cameras carry a suite of AI functions: anomaly detection, optical character recognition, and object detection and classification. These combine with deterministic, rules-based algorithms. B&R positions the hybrid as balancing AI's adaptability against the speed and repeatability of conventional vision. Product type identification, subtle defect detection and printed code verification execute in a single pass on one device, and the system switches AI models during operation. Every hardware component needs a single cable; optional hybrid connections support daisy-chain wiring, which removes external trigger sensors and cuts cabling cost.

Why lighting calibration dominates the spec sheet

Deep-learning inference inherits every weakness of its input data. A model trained under one illumination distribution misclassifies when strobe intensity or timing drifts, and vision performance has always depended fundamentally on lighting. B&R's countermeasure is a factory-calibrated lighting system that the company says improves imaging repeatability by a factor of ten or more — cleaner input data for the models, fewer false positives, more stable long-term performance. Options include illumination integrated in the camera and external synchronized modules. A dedicated Flash Controller locks light pulses to motion, automatic modulation suppresses stray light, and the design supports bright-field and dark-field strategies. Integrated strobe control eliminates additional hardware.

Synchronization without encoders

The synchronization claim rests on B&R's patented fieldbus integration, which connects vision directly to the control loop alongside motion axes, robotics and HMI communication. Trigger signals originate from the controller or the motion application. In high-speed, dynamically changing production environments, that removes the separate camera encoders such lines normally carry.

The processing gain comes from a newly developed JIT compiler that generates executable machine code when the application loads rather than interpreting it at runtime. Paired with the quad-core processor, it delivers the up-to-75-percent reduction in measurement processing time — no dedicated vision PCs required.

mapp Vision: the engineering layer

mapp Vision bundles the hardware and software inside B&R's Automation Runtime and Automation Studio environment. Control programmers implement vision tasks with minimal coding and no separate process variables. Camera images drop into HMI applications in a few clicks; parameters, lighting and trigger conditions change on the fly. Applications live on the controller, so replacing a camera preserves the data. B&R says multiple machines can link this way without sacrificing stability or inspection quality.

Anomaly detection at 60 ms

Production anomalies — color deviations, scratches, misalignment, missing components, dimensional inconsistencies — are rare events relative to standard output. That scarcity starves supervised learning of training samples. Visual Anomaly Detection (VAD) sidesteps the problem: models train on images of "good" production states only, then flag deviations from them. B&R reports detection times of approximately 60 ms. With AI processors from partners specializing in intelligent processing units, inference times drop significantly compared with conventional approaches, per the company. Training needs 30 to 100 images depending on failure complexity, simulated offline in mapp Vision with no additional hardware.

Toothbrush manufacturing shows the operating environment. CNC-driven tufting machines embed bristle bundles into rotating, turret-positioned handles at extreme synchronization levels. Roughly 60 "good" images teach the system normal production characteristics; engineers then parameterize the model for defined failure modes so it distinguishes subtle defects at line speed without slowing production.

Twelve months to full deployment

The spec sheet outpaces the rollout. Customers across multiple market sectors are trialling the system, but full deployment — model training plus machine-learning optimization — can take up to 12 months, and manufacturers often stay quiet during development to protect competitive advantage. B&R acknowledges the constraints: models must be stable, reproducible and aligned with business objectives, and deep-learning systems are computationally intensive, so integration into existing automation infrastructure demands careful engineering.

The adoption question follows directly. When inspection data becomes a control parameter rather than a post-process record, quality departments must decide how to validate a deep-learning model whose behavior changes with every retraining cycle — and whether their existing calibration and traceability practices extend to it at all.