Physical AI Enables Adaptive Welding Automation

Industrial robots have traditionally been deployed on a simple premise: define the task, program the path, constrain the environment, and repeat. That model has delivered enormous value in stable, high-volume production, but it also explains why variable processes remain difficult to automate.

Manufacturers now face skilled labor shortages, rising productivity demands, shorter product cycles, more changeovers, less predictable part presentation, and tighter quality requirements. In this environment, automation needs to perceive, learn, decide, and act in the complex conditions of real production.

What Physical Artificial Intelligence Means for Industrial Applications

During a recent conversation with experts from NVIDIA, Santa Clara, Calif.; SICK, Bloomington, Minn.; Inbolt, Paris, France; and Universal Robots (UR), Odense, Denmark, the discussion turned to what physical artificial intelligence (AI) means in practical industrial terms. The clearest answer is not “robots that think” in some abstract sense. It is robots that can handle variation without requiring manufacturers to redesign the world around them. That means fewer fixtures, fittings, and manual workarounds.

Amit Goel, director of product management for autonomous machines at NVIDIA, described physical AI as the next wave after generative and agentic AI.

“This is where AI understands the data from the real world and takes actions to control and orchestrate things in the real world,” he said. “Robotics is at the heart of that evolution of AI.”

Physical AI is not simply digital AI placed on a robot. Industrial robots must interpret vision, force, torque, touch, position, speed, and sensor inputs, then act with low latency and predictable behavior. Unlike a chatbot, a robot’s output is physical motion, with consequences for people, equipment, and product quality.

Grounded in Reality

Physical AI must be grounded in real industrial systems, not only research demonstrations. Simulation, synthetic data, and web video can help train models, but they cannot fully capture contact, friction, compliance, wear, or how a specific robot applies force. For contact-rich tasks, small variations in timing, torque, or part position can determine success.

As Goel noted, human video, web data, and simulation can support training and testing, but real robot data is needed to close the gap with the factory floor.

“The robots in the future are going to be driven by data,” he said. “We’re not going to be programming them anymore. They’re going to learn from the data.”

Physical AI can be framed around four core capabilities:

1. Perception and awareness
2. Planning and decision making
3. Adaptive execution
4. Learning and resilience

In factory terms, that means recognizing parts, replanning motion, adjusting force or speed, recovering from small errors, and improving task performance over time.

In welding, those capabilities are especially relevant because the process is sensitive to small changes in part fitup, joint location, torch angle, travel speed, and workpiece presentation. A fixed path may work on a perfectly fixtured part, but real production parts are rarely perfect. Physical AI can help the robot smooth motion, adjust execution, and use sensor feedback to keep the process within an acceptable window.

Sensing and Variability

Dominik Birkenmaier, global industry manager at SICK, highlighted AI-enabled robot guidance systems that combine image and depth data with pretrained neural networks. In variable applications such as depalletizing and bin picking, he noted that this can reduce setup, as “No training, CAD files, nor manual setup are required.”

Jan Jarvis, senior manager of visualization at SICK, added that realistic virtual sensors can help teams test systems early and explore rare scenarios safely. However, robustness is key. “AI can be extremely powerful, of course, but industrial systems must just work in the end, especially if it comes to safety,” Jarvis said.

For Inbolt’s Rudy Cohen, the same principle applies on the motion side. The manufacturer mounts a camera on a robot arm to estimate part pose in real time and update trajectories during execution. “Most robots in a factory floor are still essentially blind,” Cohen explained. “So they execute perfect trajectories in very imperfect worlds.”

That mismatch between perfect trajectories and imperfect worlds is the central problem physical AI addresses. Cohen gave an example of parts loaded into racks that were never perfectly consistent because operators and forklifts introduced small shifts. Traditional automation would misalign or stall. But with real-time trajectory correction, the robot could continue working even when the part was offset or rotated.

The same issue appears in robotic welding on high-mix parts and large fabrications. Upstream cutting, forming, tacking, and clamping introduce variation before the robot starts. Physical AI does not remove the need for good workholding or process control. Still, it can reduce dependence on perfect presentation by helping the robot compare the intended weld path with the actual joint position and adjust accordingly.

For manufacturers, the value of physical AI lies in faster commissioning, fewer fixtures, less reteaching, reduced downtime, improved quality inspection, adaptive material handling, dynamic path planning, and better performance in force-sensitive tasks.

Physical AI is most useful where variability is currently expensive, such as is welding operations that include high-mix parts, large fabrications, inconsistent fit-up, changing joint locations, and cells where excessive fixturing or reteaching has limited the business case for automation.

There are still limits. Fully end-to-end AI systems remain difficult to prove over long periods. Applications involving extreme accuracy or challenging physics are still active areas of development.

However, there is a practical path for companies considering physical AI: start with simple, bounded applications where AI can improve perception, localization, inspection, guidance, or contact behavior within a controlled safety envelope. And remember, the robot will be learning and improving every time it performs the task it has been assigned.

Conclusion

Physical AI is moving from concept to real-world production lines. But it is not replacing industrial robotics; it is extending it. Welding robots will still need good process control, safety, repeatability, and sound application engineering.

With the emergence of physical AI, they can begin to handle more of the variation that has traditionally made welding automation difficult. In that sense, physical AI is not about making robots magical. It is about making them more useful in the imperfect conditions where welding actually happens.

For more information, watch the on-demand webinar, “Discover How Physical AI is Transforming Industrial Automation” at urrobots.com/physicalAIwebinar.

ANDERS BILLESØ BECK (abbe@universal-robots.com) is vice president, AI robotics products, Universal Robots, Odense, Denmark.