Should Robots Forget Clutter After A Demo
MIT CSAIL's Masked IRL result is interesting because it treats a robot demo like a noisy room, not a perfect lesson. The reported gains are nice — up to 15% better than prior language-conditioned IRL methods, with up to 4.7x less data — but the part I care about is what the robot learns to ignore. A person may steer around a laptop because the laptop matters. Or the laptop just happened to be on the desk. Same with an elbow, a chair, a camera bag, a half-open drawer. The test I want is almost rude: teach the robot once, then move the irrelevant stuff. Put the laptop somewhere else. Take the notebook away. Change the person leaning on the table. If the robot still reaches the mug without dragging yesterday's clutter into the plan, then it learned the chore instead of the room accident. Before motion, I would make it say the quiet part: target is the mug, laptop is protected, notebook is ignored, person is outside the path. One boring teach-back screen could save a lot of trust.
Comments
The normal-person version is: don’t make me clean the room before I teach the robot. If I show it where mugs go while there’s mail on the counter and a kid’s backpack on the chair, it should ask one plain question before moving: “Is the mail part of the job, or just in the way?” That tiny interruption is less annoying than a robot quietly learning the clutter.
Yes, and the delete button matters. If the robot learns around a pile of mail during one demo, I need to be able to say: never store that pile as a rule. Some clutter is context, some is private, some is tomorrow’s problem. A good robot should leave the lesson smaller than the room.