The next phase of AI may not happen on a screen. It will happen where machines have to sense physical forces, manipulate objects, adapt to changing conditions and execute tasks reliably. Agile Robots and its Franka Robotics subsidiary used two European events in August to demonstrate how that transition is taking shape—from AI-generated robot training data to force-controlled assembly and flexible industrial welding.
Artificial intelligence has become remarkably good at generating text, images, code and increasingly complex decisions. Robotics presents a harder problem.
A physical AI system cannot simply produce a correct answer. It has to move a real object, respond to forces, compensate for variation and complete a task without damaging the equipment or the component in front of it.
That is where Agile Robots is positioning its technology.
The Munich-based robotics company and its Franka Robotics subsidiary used two August events in Germany and Switzerland to show different parts of the same physical AI stack: robot hardware, sensing, human demonstrations, training data and industrial execution.
At the all about automation event in Zurich on August 26–27, Agile Robots demonstrated the Diana 7 robotic arm for force-controlled assembly and the Thor 12 for flexible welding.
Diana 7 is designed around seven-axis motion and torque sensing in all seven joints. Those sensors allow the robot to detect forces during manipulation rather than relying exclusively on pre-programmed trajectories. Agile Robots says the system can use that feedback for applications such as inserting an engine head, where detecting abnormal resistance can help identify positioning problems and reduce the possibility of jamming or component damage.
That capability illustrates an important distinction between conventional automation and physical AI.
Traditional industrial robots are highly effective when the environment is predictable. A component arrives in a known position, the robot follows a defined path and the process repeats.
Force-controlled robotics introduces another layer of information.
Instead of asking only, “Where is the robot?”, the system can also respond to “What is the robot feeling?”
That matters in assembly tasks where small variations in component position, tolerances or contact forces can affect the outcome.
The second Zurich demonstration, Thor 12, approached the problem from another direction: making industrial robotic welding easier to configure and adapt.
The Thor family covers multiple payload classes, with the Thor 12 rated for a 12-kilogram payload and a 1,300-millimeter reach. Agile Robots lists repeatability of up to ±0.05 millimeters for the Thor 12.
The company’s demonstration focused on drag-and-drop teaching and low-code control, allowing users to create and modify welding paths without relying entirely on conventional robot programming. According to Agile Robots, Thor 12 can maintain welding performance across corner, vertical and inclined welds, including applications involving gaps as narrow as 1 millimeter.
The significance is less about one welding demonstration than about reducing the programming barrier around industrial robots.
Manufacturers increasingly need automation that can be redeployed as production requirements change. A robot that takes weeks of specialist engineering work to reconfigure is less flexible than one that operators and integrators can teach and adjust more directly.
The second August event highlighted a different requirement for physical AI: data.
At IJCAI-ECAI 2026 in Bremen, Franka Robotics demonstrated a workflow built around its GELLO Duo teleoperation system, the FR3 Duo dual-arm research platform and Franka LABS.
The setup allows a person to teleoperate the robot and demonstrate bimanual manipulation tasks. Those demonstrations can then be captured as structured data for robot-learning research.
Franka describes the FR3 Duo as a platform combining teleoperation, data collection and policy execution. The system is designed to let researchers demonstrate skills, generate datasets and subsequently deploy and benchmark learned policies on the same platform.
That connection between demonstration and deployment is becoming one of the central challenges in robotics AI.
Large language models benefited enormously from massive datasets. Robots do not have an equivalent supply of high-quality real-world manipulation data. Physical interactions are expensive to collect, difficult to standardize and often contain information that is hard to reproduce in simulation.
Human demonstrations offer one route around that problem.
Instead of manually programming every movement, researchers can demonstrate how a task should be performed and capture the resulting robot trajectories and sensor information. Those examples can then contribute to datasets used to develop policies for manipulation.
The approach also reflects a broader industry shift toward embodied or physical AI.
The International Federation of Robotics reported that 542,000 industrial robots were installed worldwide in 2024, more than twice the level a decade earlier. Asia accounted for 74% of installations, while Europe represented 16%.
The installed base creates an important opportunity for AI—but also a substantial engineering challenge.
The next generation of robots needs to operate in environments that are less structured than traditional factory automation. They need perception, force feedback, learning, adaptable software and increasingly sophisticated data pipelines.
McKinsey describes this as a transition toward physical AI, where robots can perceive environments, learn and adapt rather than simply execute narrowly defined programs. The consultancy has argued that the value opportunity extends well beyond humanoids, including manufacturing and logistics applications.
Agile Robots’ two August demonstrations therefore fit into a larger architectural picture.
Franka’s research platforms address the data and learning side. Agile Robots’ industrial systems address deployment.
The connection is important because collecting demonstrations is only useful if those datasets can eventually contribute to reliable behavior on physical machines. Likewise, an intelligent industrial robot needs more than capable hardware if manufacturers cannot efficiently teach, program and adapt it.
This is where the competitive landscape is becoming increasingly crowded.
Companies including ABB, FANUC, Yaskawa, KUKA, Universal Robots, Boston Dynamics, NVIDIA and Google DeepMind are pursuing different approaches to AI-enabled robotics. Some emphasize industrial automation, some collaborative robots, and others foundation models, simulation or general-purpose robotic intelligence.
The emerging competition is consequently moving beyond mechanical specifications.
Robot payload, reach and repeatability still matter. But so do training-data pipelines, simulation, perception, force sensing, AI models, developer tools and the ability to move from a research demonstration into production.
For manufacturers, that could eventually change how robotic automation is purchased.
Instead of evaluating a robot solely as a programmable machine, buyers may increasingly evaluate the complete physical AI platform around it: how quickly it can learn a task, how much data is required, how easily it can be redeployed and how reliably it performs when production conditions change.
For Asia-Pacific manufacturers and integrators in particular, that shift could be significant. The region already represents the largest share of global industrial robot deployment.
The next competitive advantage may therefore come from connecting AI research with industrial execution.
Agile Robots is betting that physical AI will not be defined by a single robot or model. Instead, it will emerge from a connected stack in which humans provide demonstrations, robots capture physical experience, AI systems learn from that data, and industrial platforms turn those learned capabilities into repeatable production tasks.
That is a considerably more difficult proposition than putting an AI model into a machine.
It is also where the commercial value of physical AI is likely to be tested.
Market Landscape
Physical AI is moving from tightly controlled robotics demonstrations toward systems designed to operate in more variable real-world environments.
The underlying market is already substantial. The International Federation of Robotics recorded 542,000 industrial robot installations in 2024, with annual installations exceeding 500,000 for the fourth consecutive year.
The emerging opportunity is to make that installed base more intelligent and adaptable.
Three technology layers are increasingly converging:
- Robotic hardware: arms, actuators, force sensors, cameras and end effectors.
- AI and data infrastructure: demonstrations, datasets, simulation, perception models and learned policies.
- Deployment software: programming, orchestration, monitoring and tools that allow manufacturers to adapt robots to changing workflows.
The shift also creates a strategic opening for companies that can bridge research and production.
Franka’s FR3 Duo, for example, is explicitly designed around demonstration, dataset generation and policy execution, while Agile Robots’ industrial portfolio emphasizes force-controlled manipulation and flexible automation.
That convergence is likely to become increasingly important as companies move from isolated AI robotics pilots toward systems that must operate for complete production shifts. McKinsey notes that moving general-purpose robots from pilots into productive applications will require substantially longer operating times and greater reliability.
The market is consequently evolving from robot automation toward learning-based automation.
Top Insights
- Physical AI is moving beyond robot demonstrations, combining perception, force sensing, machine learning and adaptable software to address less predictable industrial environments.
- Robot training data is becoming a strategic asset, with teleoperation providing a practical way to capture human expertise as structured demonstrations for manipulation models.
- Force control can make automation more adaptable, allowing robots such as Diana 7 to respond to physical contact rather than following position-only trajectories.
- Low-code robotics could broaden adoption, reducing dependence on specialist programming when manufacturers need to modify or redeploy automated workflows.
- The winning physical AI platforms may connect research to production, linking data collection and learned policies with industrial robots capable of executing repeatable factory tasks.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
