From Digital Intelligence to Physical Action
Physical AI is moving beyond laboratory demonstrations and into environments where robots must interact directly with machines, components, tools, and people. Agile Robots and Franka Robotics recently demonstrated this transition across two European events, showing how robotic intelligence can connect industrial automation with robot-learning data.
The demonstrations in Zurich and Bremen point to a broader change in automation engineering: the value of AI is no longer determined only by how well a system understands data. Increasingly, it depends on whether that intelligence can sense physical conditions, make decisions, and execute precise actions in the real world.
Force Control Makes Robot Assembly More Adaptive
At all about automation Zurich, Agile Robots demonstrated Diana 7 performing engine-head insertion with torque sensing across all seven joints. Instead of treating the robot as a simple position-controlled mechanism, the system monitors interaction forces during the assembly process.
This distinction matters in applications where mechanical tolerances, component variation, or small positioning errors can affect the outcome. Force feedback allows the robot to respond to the actual mechanical condition of the assembly rather than relying entirely on a predefined trajectory.
From an industrial automation perspective, this represents an important direction for robotic systems. Sensors are becoming part of the control loop, allowing robots to interpret contact conditions and adjust their behavior accordingly.
Thor 12 Targets Flexible Welding Operations
Agile Robots also presented Thor 12 for robotic welding applications. Its drag-and-drop teaching and low-code control approach is intended to simplify the creation and modification of welding paths.
The ability to adjust paths for corner, vertical, and inclined welds is particularly relevant to production environments with frequent product changes. The reported capability to handle gaps as narrow as 1 mm also illustrates how path control and process adaptation can influence welding quality.
In my view, the practical significance of systems such as Thor 12 is not simply that they make robot programming easier. The larger engineering benefit is the reduction of programming effort when production requirements change. This can make robotic automation more applicable to smaller batches and more variable manufacturing processes.
Robot Training Data Becomes Part of the Automation Stack
The Bremen demonstration presented a different part of the Physical AI equation. Franka Robotics used Franka GELLO Duo to teleoperate the FR3 Duo, while LABS captured the demonstrations as structured training data for bimanual manipulation.
This approach changes the relationship between human operators and robot programming. Instead of manually defining every robot movement through conventional programming methods, human demonstrations can provide examples from which intelligent robotic behaviors can be developed.
For AI developers, the important element is the data-generation workflow. Teleoperation provides a way to capture human manipulation behavior, while the robotic platform supplies the physical execution environment and associated data.
Bimanual Manipulation Requires More Than Motion Planning
Bimanual robotics introduces additional control requirements because two robotic arms must coordinate their movements, forces, and interaction with the manipulated object.
Training data therefore needs to represent more than individual joint positions. Demonstrations can contain information about coordinated motion and physical interaction that would be difficult to reproduce through simple scripted sequences.
This is where Physical AI differs from conventional industrial robot programming. Traditional automation generally begins with a defined task and explicitly engineered logic. Robot-learning approaches can instead use demonstrations as part of the process of developing task-specific behavior.
The Connection Between AI and Industrial Automation
The Zurich and Bremen demonstrations address different stages of the same technical chain. Diana 7 and Thor 12 focus on physical execution in industrial applications, while Franka's teleoperation and data-capture workflow addresses how robotic intelligence can be trained.
The connection between these areas is significant. A robot capable of sensing its environment still needs appropriate control strategies, while an AI model trained on demonstrations ultimately needs a physical platform capable of executing its decisions.
For industrial automation engineers, this suggests that future robotic architectures will increasingly combine sensing, motion control, AI inference, data acquisition, and conventional automation technologies rather than treating them as isolated systems.
Why Physical AI Matters for Manufacturers
Manufacturers are facing increasingly variable production requirements, shorter product lifecycles, and pressure to automate tasks that were previously difficult to standardize. Fixed robot programs remain highly effective for repetitive processes, but their limitations become more visible when the environment or product changes frequently.
Physical AI offers another approach by allowing robots to incorporate sensor feedback, demonstrations, and learned behaviors into their operation.
However, AI should not be viewed as a replacement for established automation engineering. Safety functions, deterministic control, electrical architecture, motion limits, industrial networking, and process validation remain fundamental. The most practical systems will likely combine AI-based decision-making with established control and safety layers.
From Demonstrations to Deployable Systems
The most interesting aspect of Agile Robots' recent demonstrations is the connection between development and deployment. Robot training data, intelligent control, force sensing, simplified programming, and physical execution are increasingly becoming components of one broader automation architecture.
The next engineering challenge is scalability. A successful demonstration is only the beginning. Industrial deployment requires repeatable performance, predictable cycle times, integration with existing PLC and manufacturing systems, safety validation, maintainability, and measurable production benefits.
Physical AI will therefore have to prove itself not only as an AI technology but also as an industrial control technology.
A Shift in How Robots Are Engineered
The demonstrations in Switzerland and Germany indicate that industrial robotics is moving toward a more adaptive model. Robots are increasingly expected to perceive physical conditions, learn from demonstrations, and modify their actions rather than simply repeat predefined trajectories.
My view is that the strongest opportunity lies in combining these capabilities with conventional industrial automation rather than replacing it. PLCs, motion controllers, safety systems, sensors, industrial networks, and robotic platforms can provide the deterministic foundation, while Physical AI can add a higher level of adaptive decision-making.
The result could be a new generation of automation systems in which robots are not merely programmed machines, but physical computing systems capable of sensing, learning, and acting within real production environments.
