Physical AI moves factory robots toward adaptive, variable tasks

By Sophie Lindqvist

New robotics foundation models can allow a robot to learn a task from one demonstration instead of months of programming, International Federation of Robotics President Jane Heffner said.

That claim, presented in an ASSEMBLY Robots Today episode, defines the practical direction of physical AI (AI that lets machines perceive variability and adapt) on factory floors. Heffner described the transition as real and visible daily, with applications becoming more profound as systems move beyond research.

What changes when robots leave fixed paths?

For decades, industrial robots largely followed fixed, preprogrammed routes. Physical AI changes that model by letting robots perceive variability, adjust movement and act reliably as factory conditions change. Heffner said: “For decades, robots have followed a fixed, pre-programmed path. And AI now lets them perceive variability and adapt to that and then act reliably in a changing environment.”

The International Federation of Robotics has identified AI-powered autonomy as a defining robotics trend for 2026. In practical terms, autonomy links sensing to action: a robot detects a change, modifies its behavior and continues the task rather than following an unchanged sequence.

This represents more than a speed gain for a stored program; it changes how a robot responds when a part, line or operating condition differs from the programmed case. The manufacturing value comes from handling variation, not merely repeating identical motion faster.

How does perception change factory work?

Robot perception is emerging as an enabling layer. Heffner said 3D vision, which supplies spatial information about a robot’s surroundings, can guide motion in real time. The system can locate parts and correct its path on the fly, including on moving production lines.

That capability could extend automation into high-mix production (frequent product or process variation) and less structured applications that had required frequent human intervention. Instead of a worker constantly monitoring the process, robots can increasingly adapt as conditions change.

The account does not quantify a specific deployment, plant capacity gain or labor reduction. Its technical case is that perception permits immediate path correction and adaptation can reduce repeated supervision; plant-scale conclusions still require validation against production data.

Can one demonstration replace months of programming?

Heffner identified learning from demonstration as another emerging capability. New foundation models (broad AI systems that support multiple tasks) and vision-language-action models (systems linking visual information, language and physical action) can let robots acquire a task from one example.

“They let a robot basically pick a task from a single demonstration instead of like months and months of programming, right? And it can become real almost immediately,” Heffner said. That shortens the setup sequence for variable tasks.

For manufacturing planners, the practical implication is faster task setup in variable applications. The development could reduce the programming traditionally required to deploy robots while widening the set of tasks considered suitable for automation.

The account provides no named factory deployment, production-volume figure or measured cycle-time improvement (time required to complete an operation). The one-demonstration claim therefore remains a model capability, not a documented operating result. Any plant-scale conclusion will require production data showing that the learned task repeats reliably under real conditions.

What should manufacturers watch?

The next test is whether perception and demonstration learning move from research into sustained factory use. Three outcomes would carry more weight than another laboratory demonstration: operation on variable production lines, less frequent human intervention and measurable reductions in deployment programming.

That evidence would show whether physical AI can deliver the operating change Heffner describes, rather than simply demonstrate a new model architecture. It would also clarify whether faster task setup produces dependable output when part positions, line motion or other conditions vary.

Robots Today examines current robotics developments, while ASSEMBLY’s companion series, The Future Is Robotic, takes a broader view of the technology’s direction. ASSEMBLY releases a new episode of The Future Is Robotic monthly on AssemblyMag.com. The publication also directs readers to Ask ASM, a smart AI search tool for assembly and manufacturing questions.

The next milestone is a dated production deployment reporting repeatability, cycle time and programming savings against a conventional programmed robot. The account provides no launch date or operating benchmark, so a measured factory result will carry more weight than the capability claim.