Original title: Seeing Less Is Not Seeing Safely: Privacy Leakage from Task-Scoped Robot Perception Exports
Yuqiao Xu, Erman Ayday · September 2026
Home robots see a lot of personal spaces to do their jobs. Even if a robot keeps raw camera video on the device, it still sends simplified data summaries to outside planners, cloud services, and logs. These summaries can accidentally reveal private household details through their shapes, layout, and goals. Researchers created Task-Functional Perception Distillation to test and control how much private information these exported summaries leak. They tested robots doing navigation, collision checking, and finding objects. They found that different data exports can perform tasks equally well while leaking drastically different amounts of private information. Swapping exact target labels for target regions stopped private category leaks almost entirely while keeping task success high. Making geometric shapes coarser also cut leaks, but it hurt collision performance.
Suppose you're at a library and you only jot down the titles of books you borrow instead of carrying the books themselves. But the titles alone might tell someone about your interests or secrets. Here, robots do something similar with home data, sharing only what seems necessary for tasks, but this shared information can still give away private details.
Domestic robots share processed information to function, but this can expose private home details. People using these robots need assurance that their data isn't being leaked, even in structured forms.
The abstract doesn't specify if the findings apply to real-world domestic environments beyond simulations. The evaluation is limited to specific tasks and may not generalize to all perception exports.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Seeing Less Is Not Seeing Safely: Privacy Leakage from Task-Scoped Robot Perception Exports — Yuqiao Xu, Erman Ayday
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.