Teradyne Robotics says the next phase of industrial automation depends on AI models that control a robot's motion directly, in real time, rather than the fixed, pre-programmed paths that have defined the field for decades. In an interview with Computerworld, Andrew Pether, Head of AI Partnerships at Teradyne Robotics, described the shift as "physical AI," a break from robots that historically "simply moved to fixed positions with essentially zero flexibility to handle any variation in the world around them."
Under this approach, the AI model itself makes the motion decisions, drawing on feedback from wrist- and workspace-mounted cameras, robot position data, and integrated force and torque sensors. Engineers who currently rely on teach pendants to build up fixed waypoints will care about the training method Pether described: instead of scripting logic point by point, an operator who already knows how to perform a task can demonstrate it once, and the robot learns from that demonstration. Pether said post-training and fine-tuning are still required before a robot performs reliably, and general-purpose models that adapt to new tasks without extra tuning remain a longer-term goal.
Safety is built on standards that predate the AI, Pether said. Industrial robotics has roughly 50 years of established safety standards behind it, and current systems already enforce hard constraints, such as maximum speeds and virtual boundaries, regardless of whether a human programmer or an AI model generates the motion. Universal Robots, the Teradyne Robotics group company that builds lightweight collaborative arms, has already removed physical safety cages in some deployments, relying instead on certified virtual safety limits.
Training data remains a bottleneck. Pether said video of people performing tasks by hand is useful for inferring gravity and object behavior but does not capture contact forces, so Teradyne Robotics combines real-world sensor data with photorealistic simulation. He said the company has also built an internal data-generation system it calls AI Trainer, though he did not detail its mechanics in the interview.
Pether framed manufacturing as the logical starting point for physical AI because, despite variation on the shop floor, tasks there are still well defined. He said complexity increases sharply once the technology moves into less structured settings, such as home robotics, where a system may be expected to handle almost any task.
