The industrial automation industry has been talking about AI for years. Much of that conversation has been at the outside of the machine. Cameras with embedded AI, SCADA analytics, cloud dashboards that flag anomalies. The actual control layer, the PLC, the motion controller and the realtime execution of logic where machines physically move, until now has been mostly untouched by AI.
Beckhoff stated clearly at SPS 2025 and followed this up with product releases, says that this is about to change. Their term for it is ‘physical AI’. Applying artificial intelligence directly into the controls, where it can actively change machine motion, process data, and make descisions in real time is happening.
TwinCAT Machine Learning: The Basics
Beckhoff’s machine learning capability within TwinCAT has been developing for years through the TF3800 TwinCAT Machine Learning Runtime module and the associated TwinCAT Machine Learning Creator. These tools allow trained neural networks to be deployed within the TwinCAT runtime. This means that machine learning inference can run on the same industrial PC, in the same environment, as the rest of the machine’s PLC logic.
This was a big first step. Machine learning deployments in this case pushed sensor data out to an edge server or cloud platform, run inference there, and return a result to the control system. This adds latency, relies on network dependencies, and created a system boundary that complicated safety. Running inference inside TwinCAT eliminates these issues. The Beckhoff ML model sees the same data the PLC sees, at the same cycle rate, without any communication latency.
The TwinCAT Machine Learning Creator platform is designed to make model creation accessible to automation engineers. The platform can generate efficient neural networks automatically from training data, with the goal of producing models that are optimized for the computational limits of an industrial PC rather than requiring a data center, GPU or cloud server to run.
TwinCAT CoAgent: AI Assistance Across the Engineering Lifecycle
A new feature is TwinCAT CoAgent, which addresses a different use case. AI-assisted engineering can be used via CoAgent rather than AI assisted machine operation. CoAgent is embedded directly in the TwinCAT development environment. This is the same Visual Studio based workspace where engineers write PLC code, configure I/O, and develop HMI programs.
CoAgent uses natural language prompts and responds with suggestions for PLC function blocks, I/O changes, and HMI layouts. It is contextually aware of the current project, meaning it generates code that fits the project structure and uses variable names that are already in use. CoAgent supports multiple LLM’s including OpenAI models, Anthropic models, and locally hosted ones, giving facilities with data security requirements the option to keep prompts and responses in house.
Experienced engineers can use CoAgent to accelerate the generation of usable code, sensor input handling, alarm management, PID loop setup and motion sequences. This frees their time for the design and commissioning work that truly requires expert judgment. Less experienced engineers can use it to learn unfamiliar function blocks and build starting points for new programs and tasks.
Practical Note: TwinCAT CoAgent’s code suggestions should always be reviewed by a qualified engineer before deployment. Like any AI code generation tool, it produces plausible-looking code that may contain logic errors for specific cases. Treat it as a capable first draft, not a finished product.
AI in the Control Layer: What ‘Physical AI’ Actually Means
Beckhoff’s ‘Physical AI’ concept means an AI model is running at the code and control level to directly influence the physical machine, not just for monitoring or reports. This is distinct from edge analytics or SCADA level AI, and it has some important applications.
Practical applications within TwinCAT include condition monitoring models that run inference on vibration, temperature, and current values to detect developing mechanical issues such as bearing wear, imbalances and lubrication quality in real time without routing data off the machine. These models can trigger maintenance alerts or controlled shutdowns based on the existing machine state, reducing unexpected failures and unnecessary maintenance.
Advanced cases involve models that take part in motion control loops. For example, learning models that correct for system level mechanical errors (backlash, flex, thermal drift) that are difficult to see analytically but which can be learned from observed motion data. This type of application is where the term ‘physical AI’ earns its meaning. The AI is not observing the machine, it is controlling it.
The ATRO Robot and AI-Driven Robotics
Beckhoff’s ATRO robot platform is one of the most visible demonstrations of their “physical AI”. ATRO is a modular robot architecture with arm segments and joints that can be configured into different geometries. These are controlled entirely through TwinCAT. Because the robot runs on the same platform as the rest of the machine’s automation, it can share data, coordinate motion, and participate in the same AI inference as the broader system.
This tight integration between robot control and machine PLC is something that traditional industrial robots, which are typically self contained systems communicating with the PLC over a fieldbus interface cannot offer. The data available to the AI models includes not only robot joint positions but the full process context such as upstream sensor readings, downstream inspection results, timing relative to other machine axes. This richer data environment makes better ML models possible.
Hardware Requirements: Which Beckhoff IPCs Support ML Inference?
Running “physical AI” inference in real time requires meaningful CPU resources. Beckhoff’s current IPC lineup includes multiple platforms that are designed for this workload. The C6670 high-performance rail mount IPC and the C6030 cabinet PC, both featuring Intel Core i series processors with multiple cores, are capable of running TwinCAT Machine Learning workloads alongside standard PLC and motion tasks.
For applications the need GPU accelerated inference such as larger networks and computer vision tasks, Beckhoff also offers IPCs with NVIDIA GPU options that can support CUDA based AI alongside the standard TwinCAT real time environment. TwinCAT ML is designed to transparently route inference workloads to available GPU resources when they are present.
As TwinCAT PLC++ delivers improved CPU efficiency on existing hardware, the compute headroom available for ML inference on a given IPC has increased. Another valuable positive between the two major platform developments Beckhoff shipped in 2025.
Data Infrastructure: ADS, OPC UA, and Analytics Integration
Running machine learning in TwinCAT is only useful if training data can be collected efficiently and models can be updated as the machine conditions evolve. Beckhoff addresses this through TwinCAT’s existing data infrastructure. The ADS (Automation Device Specification) protocol for internal data access, OPC UA for external connectivity, and TwinCAT Analytics for historical data logging and analysis.
TwinCAT Analytics allows high frequency data from any TwinCAT system to be logged, streamed, and analyzed either locally or in the cloud. This data pipeline is what makes model training and retraining practical. Engineers can capture labelled training data (normal operation, fault conditions, process variations) from production machines, retrain their models offline, and deploy updated models back to the TwinCAT runtime without taking the machine offline.
Where Beckhoff’s AI Strategy Fits in the Broader Industrial Landscape
Beckhoff is not the only automation vendor pursuing AI integration in the control layer, but they are among the first to implement this successfully. Their advantage is that TwinCAT is a genuinely open, PC-based platform with the compute resources and software architecture to run serious workloads. This is something that embedded PLCs from traditional vendors simply cannot match without significant additional hardware.
Siemens is pursuing similar directions through their Industrial Edge platform and the SIMATIC S7-1500 TM NPU (Neural Processing Unit) module. Rockwell Automation has been building analytics capabilities through FactoryTalk. But neither platform offers the same degree of integration between ML inference and the core real time control layer that TwinCAT enables. Precisely because TwinCAT runs on a general purpose PC operating system with direct access to modern processor capabilities.
For equipment builders or integrators evaluating where to invest in AI capable automation infrastructure, the Beckhoff TwinCAT platform with TwinCAT PLC++ and CoAgent represents a mature, deployable, and continuously evolving option.
Getting Started with TwinCAT Machine Learning
Beckhoff users who want to begin exploring TwinCAT ML capabilities should look at the TF3800 TwinCAT 3 Machine Learning Runtime first, which is available as a licensed module for TwinCAT 3 and TwinCAT PLC++. Beckhoff provides sample projects and documentation through their InfoSys platform.
A practical first project for most teams is a condition monitoring application on a motor or drive axis. Collecting vibration or current data, training a simple anomaly detection model, and deploying it to a test system to validate the data pipeline before applying it to production equipment. This gives development teams hands on experience with the full workflow of data collection, model training, TwinCAT deployment, and runtime integration with manageable risk.
Beckhoff SEO Article Series|Celtex Electric & Automation|www.celtex.ca/beckhoff