From AI Concepts to Physical Robot Execution
Physical AI is moving beyond demonstrations in software environments and into machines that must interact directly with real components, tools, surfaces, and production processes. Recent demonstrations by Agile Robots and Franka Robotics show this transition from two complementary perspectives: industrial robot deployment and the collection of data required to train intelligent robotic systems.
The demonstrations in Zurich and Bremen indicate that the next stage of robotic automation will depend not only on AI models, but also on sensing, motion control, force feedback, human-machine interaction, and structured training data. In practical manufacturing environments, these elements have to operate together before AI can deliver repeatable physical results.
Diana 7 Uses Joint Torque Feedback for Assembly
At all about automation in Zurich, Agile Robots demonstrated the Diana 7 robot performing force-controlled assembly. The seven-axis robot uses torque sensing at its joints to monitor insertion forces during an engine-head assembly operation.
This approach changes the control problem from simply moving a robot to a predefined coordinate toward continuously evaluating the physical interaction between the robot and the component.
Real-time force information can help the robot determine whether an insertion is progressing correctly and identify abnormal resistance. From an automation engineering perspective, this is particularly relevant for assembly operations where positional accuracy alone is insufficient. Small variations in component position, tolerances, or mechanical alignment can otherwise produce excessive forces, jamming, or component damage.
The broader significance is that force feedback provides another control variable alongside position and velocity. This gives robotic systems a better basis for handling processes where mechanical contact is unavoidable.
Thor 12 Targets Flexible Robotic Welding
Agile Robots also presented the Thor 12 for robotic welding applications. The system uses drag-and-drop teaching and low-code programming to simplify the creation and modification of welding paths.
The demonstrated system is designed to handle corner, vertical, and inclined welds, including welding situations involving gaps as narrow as 1 mm.
For manufacturers operating high-mix production lines, flexible programming can be as important as robot payload or motion speed. Traditional robotic welding often requires considerable programming effort whenever part geometry or production requirements change. A more accessible programming workflow can reduce the engineering work associated with modifying robot paths and make automation more practical for smaller production batches.
The important point is not simply that a robot can weld. Modern industrial robotics increasingly needs to accommodate production variation without requiring a complete re-engineering of the automation cell.
Robot Training Data Becomes Part of the Automation Architecture
The Bremen demonstration at IJCAI-ECAI 2026 addressed a different part of the Physical AI problem: how to generate useful data for training robot behaviors.
Franka Robotics demonstrated a workflow in which operators used the Franka GELLO Duo to teleoperate the FR3 Duo. LABS then captured these demonstrations as training data for bimanual manipulation.
This creates a direct connection between human physical expertise and machine-learning datasets. Instead of relying exclusively on manually programmed trajectories, developers can record how a human performs a manipulation task and convert those demonstrations into structured data for robot-learning applications.
For industrial automation, this is an important development because programming every possible physical interaction manually becomes increasingly difficult as robots move into more variable tasks.
Teleoperation Can Reduce the Data Collection Bottleneck
Robot learning requires large quantities of useful physical data, but collecting that data is more complicated than collecting conventional digital datasets. A robot must physically interact with objects, and the resulting data needs to describe meaningful actions, positions, forces, and task outcomes.
Teleoperation provides one practical method for generating this information. Human operators can demonstrate behaviors that would be difficult to describe using conventional robot programming methods, while the robotic system records the corresponding movements and interactions.
In my view, this could become one of the more important links between traditional industrial robotics and AI-based automation. The value of teleoperation is not limited to remote control; it can also function as a mechanism for transferring human process knowledge into machine-readable training datasets.
The Industrial Value of Physical AI Depends on Integration
The Zurich and Bremen demonstrations should not be viewed as two unrelated robotics showcases. They represent two sides of the same technical development.
On one side, industrial robots require increasingly sophisticated sensing and control to interact with real production environments. On the other, AI systems require physical data to learn how those interactions should be performed.
Force-controlled assembly provides physical feedback during execution. Teleoperation and data capture provide examples from which future robot behaviors can be learned. These capabilities can eventually form a closed development cycle in which robots collect data, AI models learn from that data, and improved models are deployed back into physical automation systems.
From Conventional Programming to Data-Driven Automation
Conventional industrial automation remains highly dependent on deterministic logic, fixed trajectories, predefined process parameters, and structured production conditions. These methods remain appropriate for many high-volume and repeatable applications.
Physical AI is more relevant where variability becomes difficult to manage through fixed programming alone.
The combination of force sensing, flexible programming, teleoperation, and machine-learning datasets suggests a gradual transition rather than an immediate replacement of conventional automation. PLCs, robot controllers, sensors, safety systems, and industrial networks will continue to provide the deterministic infrastructure, while AI-based functions can increasingly handle perception, adaptation, and learned behavior.
A Practical Path Toward More Adaptive Robotics
The most significant aspect of Agile Robots' demonstrations is therefore not a single robot model or software feature. It is the movement toward an automation architecture in which sensing, control, human expertise, and AI training data are connected.
For manufacturers and system integrators, this approach could make automation more adaptable to variable assembly, welding, manipulation, and other tasks that are difficult to fully standardize.
The engineering challenge will be ensuring that AI-driven behavior can still meet the requirements of industrial safety, repeatability, traceability, cycle time, and process validation. Physical AI becomes commercially meaningful when it can satisfy those constraints inside an actual production system, rather than only performing successfully in a controlled demonstration.
