Industrial AI, Robotics and the Future of Factory Automation
Early applications of artificial intelligence in industrial environments consisted mostly of predictive maintenance algorithms, generative copilots and constrained pilots for data analysis. Now, manufacturing companies are contemplating a deeper transformation by integrating AI directly into physical systems.
In a recent episode of the Future Ready Podcast, host Spencer Acain sat down with Dr. Matthias Loskyll to examine this transition. Their discussion explored the emergence of physical AI, the critical relationship between virtual simulation and real-world robotics, the data architectures required to anchor intelligent applications, and the eventual transition toward autonomous, agentic workflows across the shop floor. Listen to the conversation below, or keep reading for a summary of the discussion.
Physical AI Translates Prediction into Action
The emergence of Physical AI is representative of the paradigm shift in the manufacturing sector’s expectations for AI. Mainstream AI models are primarily known for generating text and images, synthesizing data and other analytical tasks. In an industrial context, however, Loskyll pointed out that intelligence ultimately translates into physical action. Physical AI thus focuses on the integration of AI directly into physical robotics and industrial systems.
Physical AI combines sensory perception, contextual reasoning, and mechanical actuation. Rather than simply evaluating digital inputs, these systems operate within complex and dynamic physical environments. Industrial plants and factories represent severe challenges for computational systems. In the real world these systems must account for manufacturing tolerances, variations in lighting, surface and material properties and, of course, human workers that share the workspace.
Because of this inherent unpredictability, physical AI must deliver exceptionally high standards of functional safety, robustness, and auditability. Unlike consumer algorithms, where an occasional erroneous output carries low consequence, an uncalculated motion in an industrial cell can damage multi-million-dollar capital equipment or endanger human operators.
Loskyll provided an example of a physical AI system already at work. Siemens Simatic Robot Pick AI is designed to enable industrial robots to identify and pick objects it has never seen before. The system is pre-trained and can react to arbitrary shapes, textures, orientations and packaging, even identifying the best pick point for the object presented.
How the Digital Twin Fuels Robot Foundation Models
Deploying physical AI at scale introduces significant machine learning hurdles, primarily concerning training data acquisition. Capturing every potential corner case, lighting condition, and physical disruption on an active production line is not cost-effective or safe.
Instead, companies can leverage the Digital Twin and high-fidelity simulation to aid in the development of physical AI systems.
Loskyll explained how simulation environments allow engineers to generate massive volumes of synthetic data before physical robots or autonomous mobile platforms are deployed in a factory. This approach is particularly critical for the emergence of robot foundation models, including those designed for humanoid platforms and multi-axis manipulators. Simulation and the Digital Twin enables engineers to examine hundreds of thousands of iterations, uncovering issues in the virtual world.
Loskyll continued, describing how real world data can be incorporated to fine-tune foundation model behavior and make it ready for real world deployment. Pre-training with synthetic data and the Digital Twin followed by validation with real-world scenarios and hardware ensures that models meet industrial reliability standards while dramatically shortening physical commissioning cycles.
Unlocking the Industrial Data Fabric
Despite rapid advances in algorithmic capabilities, fragmented data systems continue to hold back many industrial AI implementations. Most manufacturing facilities operate within “brownfield” environments where machinery, controllers and other infrastructure have been installed, modified, and maintained over decades.

Loskyll illustrated how critical operational knowledge is rarely consolidated in manufacturing environments, and especially brownfield settings. Instead, it is scattered across programmable logic controllers (PLCs), enterprise resource planning (ERP) systems, spreadsheets and unstructured paper documentation. Furthermore, significant portions of historical process knowledge remain entirely non-digitized and may only exist in the heads of experienced technicians.
To extract tangible value from machine learning models, enterprises must establish an industrial data fabric. Rather than pursuing massive, risky data migration projects that attempt to rip and replace operational systems, a modern data fabric connects and contextualizes information where it currently resides.
This architecture relies on an underlying semantic layer, frequently implemented as a knowledge graph. This semantic layer maps the functional relationships between time-series sensor data from the shop floor, engineering specifications from the design phase, maintenance logs, and broader operational metrics. By establishing a unified, contextualized data foundation, enterprises eliminate silos, providing AI models with a reliable, verified, single source of truth.
Agentic AI and Autonomous Workflows
Loskyll then explored how maturing data fabrics and algorithmic capabilities are allowing industrial automation to move to agentic AI.
While conventional generative AI operates in a reactive, conversational question-and-answer format, agentic systems are designed to operate autonomously based on high-level objectives. An agentic system accepts an overarching goal, evaluates the current state of the environment, decomposes the goal into specific subtasks and executes those tasks across multi-step workflows.
In practice, industrial environments will not rely on a single, monolithic model. Instead, architecture will favor a multi-agent ecosystem with specialized domain agents executing specific tasks and orchestration agents managing workflows. But even these systems will continue to operate with human oversight to prevent costly or potentially dangerous mistakes from occurring.
Bringing Intelligence to the Factory Floor
Industrial AI holds immense potential to transform industrial operations, both virtually and in the real world. Successful implementations will require companies to deploy the Digital Twin for design and validation, a unified industrial data fabric to unlock legacy know-how, and agentic workflows operating with human oversight. As these pillars converge, manufacturers are moving past incremental AI pilots toward an interconnected ecosystem where physical AI and autonomous systems drive the next era of industrial resilience and productivity.
For more conversations on exciting industrial technologies featuring key experts, follow the Future Ready podcast.
Siemens Digital Industries Software helps organizations of all sizes digitally transform using software, hardware and services from the Siemens Xcelerator business platform. Siemens’ software and the comprehensive digital twin enable companies to optimize their design, engineering and manufacturing processes to turn today’s ideas into the sustainable products of the future. From chips to entire systems, from product to process, across all industries. Siemens Digital Industries Software – Accelerating transformation.