VISION 2026: Filters, SWIR and 200 GigE Define the Machine-Vision Stack

GigE Vision cameras have scaled from 1 Gigabit to 100 Gigabit and now 200 Gigabit, and that bandwidth figure frames the engineering problem behind every autonomous machine-vision system shown at VISION 2026. Three interviews recorded at the Stuttgart trade fair — with Chroma, Emergent Vision Technologies and Pleora Technologies — cover the full pipeline: optical filters and SWIR imaging at the front end, high-speed transport in the middle, and edge AI processing at the point of action.

Why does the pipeline start with optics, not AI?

Discussion of autonomous machine vision usually begins with the model: how it recognizes a defect, locates a part, or guides a robot. Georg Draude of Chroma, speaking with Vision Systems Design, argues that every one of those decisions depends on information captured upstream.

Optical filters control which wavelengths reach the sensor. Short-wave infrared (SWIR) imaging reveals information that conventional visible-light imaging cannot. Draude's core point is a constraint that no amount of software removes: a model cannot recover information the imaging system never captured. The right combination of illumination, optics, filters and camera sensor sensitivity makes relevant features more distinguishable, giving downstream software better data to work with.

That framing moves optics inside the AI system rather than treating it as a component selected after the camera and model. As specialized sensors and imaging approaches become more common, the front-end selection problem grows more acute for system designers.

What does 200 GigE transport actually require?

Eric Petz, head of marketing at Emergent Vision Technologies, discussed the evolution of high-bandwidth GigE Vision cameras, including 100 GigE and 200 GigE models, and the infrastructure needed to support data-intensive AI workloads.

Higher bandwidth lets engineers balance resolution against frame rate, but the camera is only one element. Ethernet infrastructure, network interface cards, software architecture and data paths to GPU processing together determine whether a system can handle its image stream. Petz cited GPU-direct processing as one method for moving data through the computing pipeline without bottlenecking.

He also pointed to emerging sensors — SWIR and line scan among them — and expanding lens and camera options. AI techniques such as foundation models and synthetic data may further support more autonomous manufacturing. Named applications include:

• defect detection
• bin picking
• PCB inspection
• markerless motion capture

Different sensors and imaging modes generate different data profiles, and the network and processing architecture must match the application's needs rather than the reverse.

How does perception become action?

James Falconer, product manager at Pleora Technologies, described eBUS Edge and the challenge of joining image data and AI processing at the edge. Edge processing lets systems decide close to where data is captured, and eBUS Edge supports developers working with trained or pre-trained AI models while streamlining image data movement across the network.

For autonomous manufacturing, speed alone does not define performance. Systems may need to synchronize information from multiple sensor types — 1D, 2D and 3D — so robots respond to a complete picture of their surroundings. Falconer also raised cybersecurity, including requirements tied to the EU Cyber Resilience Act, which connected vision systems shipping into Europe will have to meet.

What should engineers evaluate?

The reported takeaway from all three conversations is that the whole pipeline deserves evaluation, stage by stage:

• Does the imaging setup reveal the feature of interest?
• Can the network sustain the required throughput?
• Can processing keep up?
• Are sensor streams synchronized well enough for a dependable response?
• Are cybersecurity requirements addressed as systems become more connected?

Each stage introduces design and integration decisions that compound downstream. A 200 GigE link cannot compensate for an optical front end that never captured the feature; an edge model cannot act on data that arrived late or unsynchronized.

The open question for adopters is where the binding constraint sits in their own stack — and whether vendors' claims about end-to-end performance, most of them presented at VISION 2026 without published test conditions, hold up under measurement on the factory floor.