A word from the CiA Managing Director

Embedded AI (artificial intelligence) is the main topic on the CiA stand (hall 5, booth 410) at the SPS trade show in Nuremberg (November 25 to 27). Google AI describes embedded AI as the integration of AI into physical devices and systems, allowing them to process data, make decisions, and act locally without relying on cloud connectivity. This is achieved by running "lean" AI models on efficient hardware, leading to benefits like improved efficiency, precision, and real-time performance. Key applications include autonomous robots, smart sensors, and intelligent systems in sectors like healthcare and manufacturing.”

Embedded AI is the next logical step in improving embedded control networks. AI was first used in large remote computer farms eating a lot of energy. Of course, some of such AI applications were also fed by sensors located in front-end embedded networks. The remote AI systems processed them and provided AI-generated commands to the actuators in the front-end networks. In some embedded control applications, edge AI controllers were implemented to overcome the dependency on the central AI computing resources, requiring web-based connections, which were often to slow for real-time control systems. As said, the next logical step is the introduction of AI-based sensors and actuators, optionally connected to a local edge AI controller.

It is usual that in early days of a new technology centralized control architectures need to be applied. Normally, the next step is decentralization. The benefit is that pre-processed data can be transmitted via low-bandwidth networks in real-time. This is an advantage for all autonomously operated systems, such as AGVs (automated guided vehicles), AMRs (autonomous mobile robots), drones, driverless commercial vehicles (in agriculture, forestry, earthmoving, construction, and other industries), etc. Historically, most of them use CAN-based networks. By some means, this applies also to medical devices and laboratory equipment. There is also another positive side effect: Such embedded AI networks do not need necessarily a remote interface, which makes it easier to protect them against cyberattacks.

First embedded AI devices could be sensor-fusion devices for navigation and condition-monitoring devices, providing artificial intelligent (“smart”) pre-processed data. On the actuator side, you can think about self-configuring and self-optimizing electrical drives. Additionally, pre-emptive maintenance and AI-based condition monitoring of motors are of interest. Of course, you can use embedded real-time AI host controllers, but the direct integration of AI software into actuator devices unburdens the system designer from AI programming.

I am sure that chipmakers will soon provide appropriate AI-capable hardware consuming less power than today – for reasonable prices, of course. This enables embedded AI also for battery-powered applications. I think, CAN-based networks are ideal to connect AI-based sensors, actuators, and host controllers. The robustness and reliability of CAN technology as well as its scalability are key features for real-time control systems, featuring embedded AI.

A very first example of an embedded AI suitable CANopen profile is the CiA 462 profile for item detection devices. It specifies independent of the used sensor technology (camera, radar, ultrasonic, etc.) the process data mapped into PDO (process data object) messages. Such sensor-fusion devices could provide process data generated by embedded AI algorithms. Other CiA profiles need to be extended regarding condition monitoring parameters.

By the way, CiA distributes a flyer about embedded AI and CANopen on the SPS exhibition, and CiA representatives are ready to discuss the embedded AI topic with you on CiA’s booth. You are also welcome to schedule a meeting with CiA on the fairground in advance.