The Robot That Learned to Pour From a Teapot It Had Never Seen

When researchers at Northeastern University, Brown University, and the Boston Dynamics AI Institute wanted to test whether their robot could truly generalize a skill to an object it had never encountered, they reached for a teapot. It was not in ShapeNet, the standard shape library that underpins most robot training benchmarks. It had never appeared in a robotic demonstration. And when the robot poured from it into a mug, adapting a single human demonstration to the novel geometry on the fly, it worked. The moment is small. The implication is not. Most robot learning papers test on variations of objects the system has already seen. This one did the opposite: it deliberately chose something outside the training distribution and asked whether the skill could transfer anyway. The answer, according to a paper accepted at ICRA 2026, was yes. The method decomposes objects into semantic parts — a handle, a spout, a body — rather than treating each object as a monolithic shape to be matched. Interaction points are extracted from demonstration objects and mapped onto novel ones by matching their constituent parts. The researchers call it one-shot cross-geometry skill transfer. What it means in practice is that a single demonstration, on a single object, can generalize to a range of different objects without retraining. The approach was validated in both simulation and on a real robot arm, which used four RGB-D cameras to perceive its environment and a segmentation model to isolate individual objects. Tasks included placing a mug on a rack, stacking a bowl on a mug, and the pre-pour alignment from a teapot into a receiving mug. The teapot was the stress test. Its shape bears little resemblance to the mugs and bowls in the training set. If the method could not transfer the pour to a teapot, the researchers would have to concede that their generalization was category-level at best. It transferred. The authors are Skye Thompson, a PhD student at Northeastern; Ondrej Biza, who earned his PhD at Northeastern and now works at the Boston Dynamics AI Institute; and George Konidaris at Brown. The prior work closest to this is the same team's earlier paper on interaction warping, presented at CoRL 2023. This version is more ambitious: where interaction warping operated within object categories, part decomposition is meant to bridge across them.