The short version: MathWorks released MATLAB and Simulink R2026b on September 24, 2026. The AI pieces engineers will notice are a new Visual Inspection Toolbox for building defect detection apps, Copilot-written explanations in the Simulink Profiler, and the ability to run PyTorch models alongside MATLAB and Simulink.
What’s new
The release adds three new products: Simulink Variant Manager for handling product variants, Polyspace Test Server for automating software testing and code coverage, and the Visual Inspection Toolbox. The toolbox combines AI with image processing so teams can build and deploy visual inspection apps on factory floors and in other industrial settings.
- Simulink Copilot in the Profiler. When a simulation runs slowly, the Profiler now uses Simulink Copilot to explain the results and summarize solver behavior, which shortens the hunt for bottlenecks.
- PyTorch co-execution. The Deep Learning Toolbox can run PyTorch models alongside MATLAB and Simulink, plus a new Time Series Modeler app for forecasting and regression models.
- Point Cloud Toolbox. Supports perception work with lidar, stereo and RGB-D cameras, and mmWave radar, including SLAM and structure-from-motion workflows used in robotics and automation.
- Quality engineering apps. The statistics toolbox adds an interactive design of experiments (DOE) app and a Gage R&R app for measurement system studies.
How it builds on earlier releases
R2026b follows R2026a in April, which put Simulink Copilot and Polyspace Copilot inside the tools engineers already use and connected MATLAB to AI coding agents through the MATLAB MCP Core Server and MATLAB Agentic Toolkit. MathWorks has said its copilots read the model structure and its own documentation rather than acting as a generic chatbot. The direction is clear: AI help is being built into model-based design, testing and verification, not bolted on.
What changes in the day-to-day work
Automated visual inspection has been around for years, but building it usually meant buying a dedicated vision system or hiring someone who knew deep learning frameworks. Putting AI defect detection into a MATLAB toolbox changes who can do that work. A controls or test engineer who already uses MATLAB can collect images from the line, train a model to spot scratches, voids or missing parts, and deploy it without leaving the tools they know.
- Quality and test engineers can prototype an inspection station themselves, then pair the results with the new DOE and Gage R&R apps to prove the measurement system holds up.
- Controls and embedded engineers working in Simulink get faster answers when a model runs slowly, because Copilot explains the Profiler results instead of leaving them to read raw solver data.
- Teams with data scientists no longer have to rebuild PyTorch models in MATLAB. Running them side by side shortens the trip from a trained model to a simulated or deployed system.
- Robotics and automation teams get broader point cloud support for lidar, depth cameras and radar, which feeds bin picking, guided vehicles and other perception work.
The catch is validation. An inspection model that flags too many good parts stops a line, and one that misses defects ships them. Engineers still need to know how to build a labeled image set, measure false accept and false reject rates, and decide when a model is good enough for production.
Who should pay attention
Manufacturers adding inspection to existing lines, automotive and aerospace suppliers using model-based design, medical device makers with heavy test and verification demands, and robotics teams building perception systems. Anyone already licensed for MATLAB and Simulink can try the new features in R2026b.
What to watch next
MathWorks now ships two releases a year, and each of the last three has added more AI. R2025b brought MATLAB Copilot, R2026a added Simulink Copilot, Polyspace Copilot and a way for outside AI coding agents to drive MATLAB, and R2026b adds AI inspection and Copilot help inside the Profiler. Expect the next release to keep pushing copilots deeper into verification and testing, the part of engineering where mistakes cost the most.
For employers, the practical question is less about which features exist and more about whether the team uses them. Engineers who already work in MATLAB and Simulink every day are the ones who will turn these tools into faster test cycles and fewer escapes, so it is worth asking candidates which recent features they have actually tried.
What it means for hiring
Machine vision used to need a specialist. Toolboxes like this put AI inspection within reach of controls, test and quality engineers who already know MATLAB.
- Controls and test engineers with model-based design experience can now take on vision and inspection projects. Ask whether they have trained or validated an inspection model.
- Quality engineers who can run DOE and Gage R&R studies and also read model output are a strong fit for plants adding AI inspection.
- Copilot help speeds up debugging, but candidates still need to explain why a solver or model behaves the way it does.
Hiring for automation or embedded work? See our controls engineering recruiters and electrical engineering recruiters pages.
Interview questions for controls, test and quality engineers
- Walk me through how you would build and validate an AI inspection check for a new defect type.
- How do you decide whether a vision model’s false reject rate is acceptable for production?
- Describe a Simulink model you sped up. What was slowing it down and how did you find it?
- Have you run a Gage R&R study? What did it tell you and what did you change?
Frequently asked questions
When was MATLAB R2026b released?
MathWorks announced Release 2026b of MATLAB and Simulink on September 24, 2026.
What is the Visual Inspection Toolbox?
It is a new MATLAB product in R2026b that combines AI and image processing for building and deploying visual inspection applications in manufacturing and other industrial settings.
Do engineers need deep learning experience to use it?
Less than before, since the workflow lives inside MATLAB. But anyone putting a model into production still needs to understand training data, test sets and error rates well enough to trust the results.
More stories like this are on our AI for Engineers hub.
