See how Digi International and STMicroelectronics bring on-device AI inferencing and secure fleet management to industrial and robotics designs with the STM32MP2 processor series and Digi ConnectCore MP2.
Modern industrial and robotics systems are expected to run real-time computer vision, make autonomous decisions, and control motion, all in compact, power-efficient form factors. In this webinar, Digi International and STMicroelectronics explain how the STM32MP2 processor series and the Digi ConnectCore MP2 system-on-module bring AI inferencing and computer vision directly onto the device, without depending on the cloud or exposing sensitive operational data.
Presenters walk through a live demonstration of a robotic hand application, built on Digi ConnectCore MP2, that maps hand movement in real time and translates it into precise motor control. The session also covers how Digi TrustFence, ConnectCore Security Services, and ConnectCore Cloud Services support secure over-the-air updates, remote fleet management, and compliance with regulations such as the EU Cyber Resilience Act.
Connect with Digi
Want to learn more about how Digi can help you? Here are some next steps:
Q&A: Secure by Design, Intelligent by Default: A New Approach to AI-Powered Industrial and Robotics Systems
After the presentation, Mike Hartmann, Bob Blumenscheid and Colin Bell took audience questions covering board availability, evaluation resources, AI use cases, and guidance on choosing between the STM32MP2 and STM32N6:
When will the industrial boards that Colin presented be available?
Bob Blumenscheid: ST is building the boards with the ConnectCore MP257 modules on them, and Digi expects to have them in Q4. If you're interested in that application, whether for robotics or PLC designs, reach out to Bob to get connected with the boards.
What's the best way to get started evaluating the MP2 platform, and where can we find more information?
Mike Hartmann: The ConnectCore platforms are easy to get from DigiKey and Mouser. If you'd rather start with an ST board, the STM32MP257F-DK is the simplest option and is also available through DigiKey or Mouser. To get started, go to wiki.st.com and look under the MPU section, which has getting-started guides for OpenSTLinux.
Bob Blumenscheid: Digi's development boards come with a built-in getting-started process that downloads the current software and walks you through the document portal to begin development.
What are the typical AI applications for the MP2?
Bob Blumenscheid: The most common are vision applications, such as visual inspection and detecting people or objects, along with machine monitoring for predictive maintenance.
Mike Hartmann: Digi and ST have also seen interest in audio applications, such as sound recognition for security use cases like detecting breaking glass or gunshots. Because the MP2 can run AI workloads on the NPU, the GPU, or the Cortex-A and Cortex-M cores, developers can choose whichever processing core best fits their application.
Can you elaborate on the AI functionality possible on the NPU and GPU? Is it limited to vision, or can you run an LLM with an open source, open weight model for local inference?
Mike Hartmann: Running an LLM isn't realistic on this hardware. At a maximum of 1.35 TOPS, it would be extremely slow, so that use case is better suited to higher-end devices or the cloud. That said, the MP2 goes beyond vision: customers and partners have built applications that use more than just machine vision, taking advantage of the flexible NPU and GPU architecture. If the NPU isn't the right fit for a given workload, you can also write algorithms for the GPU using OpenCL or Vulkan.
What version of the MP2 are you using on the PLC board?
Colin Bell: It varies by application. Digi and ST typically switch between the MP255 and MP257: the MP257 is used more often when switching is needed for a zonal processor, while the MP255 is typically used for an end node. The right choice depends on the application, and the teams work with customers directly to identify the best fit.
The robotic hand video appeared to use an STM32N6 for the camera and the MP2 for the PLC side, but the MP2 also has a camera interface. When should you use an STM32N6 instead of an MP2?
Mike Hartmann: The N6 is ST's microcontroller with a built-in NPU and a MIPI CSI camera interface, and it shares the same camera and image-processing IP as the MP2. The choice comes down to use case and preference. If you want image capture and machine vision handled entirely within a single real-time microcontroller, use the N6. If you want that imaging and neural-processing capability accessible within a Linux domain, along with additional functionality, use the MP2. The two can also work together: in a humanoid robot, for example, the N6 could handle local image capture and recognition, then connect to the MP2 for further processing or to use the MP2's networking and integrated TSN switch as a bridge to the rest of the robot.
Is there detailed documentation on the NPU and GPU capabilities, specifically a diagram of how they're wired and how they handle data flows, to understand what can be run for AI workloads?
Mike Hartmann: Start with the MP2 reference manual rather than the datasheet, since the datasheet doesn't go into this level of detail. The NPU and GPU share a flexible IP: you can dedicate more processing power to the GPU, more to the NPU, or split it between both, and set priority when running both at once, such as prioritizing AI processing over display rendering. The wiki at wiki.st.com may have additional guidance, and the online support portal provides access to further technical resources.