Robots Can Now Learn to Wipe 'Gently' or 'Quickly' From Watching You

Have you ever shown someone how to do a task — holding their hand to guide the motion? That's roughly how modern robots learn. Researchers record a human doing a job, then the robot copies it. The catch: the robot copies one version only. It wipes the board at exactly the speed and pressure the human used, every single time. If you want it slower, or gentler on a delicate surface, you're stuck.

A team at Japan's Kyoto-area labs (Takumi Kobayashi, Masato Kobayashi and Yuki Uranishi) built a fix called Bi-MoDe. Their robot learns by watching a human, and it records two things at once: where the hand goes, and how hard it pushes. That second part matters because wiping, polishing, or screwing something in is all about force, not just position. Then they added something like a settings dial. Before the robot moves, you give it a simple instruction — fast or slow, light touch or firm press. The robot adjusts its whole motion to match.

Why is that hard? Because 'be gentle' isn't a number. The team solved it by feeding that instruction into every layer of the robot's decision-making brain (a Transformer — the same family of AI that powers chatbots), so the directive shapes the entire action rather than one small part of it. They tested it on a real whiteboard-wiping task, mixing speed and pressure settings, and it followed the pressure instructions better than the standard approach while keeping speed control about the same.

The honest limits: this is one task, one robot, one lab. No product, no price, no timeline. But the idea is the interesting part. Today, changing how a robot behaves usually means an engineer and a new training run. Tomorrow, it might mean saying 'easy there' — the same way you'd talk to a new hire on their first day.