Cognex has launched the In-Sight 6900 Vision Controller, a modular machine vision platform powered by NVIDIA Jetson technology that the company says handles demanding AI workloads at the edge "without the cost and complexity of PC-based systems."
The controller delivers up to 157 TOPS of AI performance, enough to run multiple high-resolution AI models simultaneously. Its modular design supports interchangeable cameras, lenses, and industrial lighting from Cognex's LOCA (lenses, optics, cameras, and accessories) portfolio, letting engineers configure hardware around the inspection rather than around a fixed-configuration box. On-device storage reaches 256 GB, and maximum sensor resolution tops out at 65 MP, wide enough to capture subtle defects across large fields of view without sacrificing detail on small parts.
What's new here is the AI tooling. The In-Sight 6900 introduces transformer-based few-sample classification that trains on as few as 10 to 20 images, a meaningful reduction from the labeled datasets typically required to build a reliable vision model from scratch. A pixel-level segmentation mode handles parts with noisy or low-contrast backgrounds where conventional rule-based tools struggle. NVIDIA TensorRT integration keeps inference latency synchronized with high-speed production timing.
The controller connects to OneVision, Cognex's collaborative AI development platform, which allows engineering teams to build, refine, and push inspection applications across multiple sites from a single environment. That architecture becomes relevant for manufacturers running the same inspection across several plants, where centrally trained models can be adapted locally and deployed to any compatible Cognex device.
Cognex positioned the In-Sight 6900 directly against PC-based vision systems, a setup still common in high-complexity inspection lines where dedicated industrial computers offload processing from embedded cameras. Edge AI controllers move that compute burden onto the device itself, reducing cabling, floor-space requirements, and potential failure points in the system. Availability and pricing were not disclosed in the announcement.
