Researchers led by Adrian Röfer have presented Cleanly-DINO, a way for robots to learn tasks such as cleaning from a few demonstrations, according to a paper accepted at a CoRL 2026 workshop and posted on arXiv. Cleaning is defined by how a surface looks rather than by where objects are, which makes the task harder to describe for a robot, the authors write.
Given a mask of the area that matters and features from a frozen vision transformer, Cleanly-DINO fits a projection that isolates the task and a measure of progress. The team trained it on varied cleaning examples, tested how well it finds unfinished areas and carries over to new visual settings, and used it as an observation encoder and reward in reinforcement learning.
They also used the model in a real-world robotic wiping pipeline to guide the task.



