Your robot can see your couch. A new tool wants it to stop seeing which room it's in.

When a home robot scans a room, the same 3D map that helps it avoid the coffee table can also tell a cloud server which room it is in and, increasingly, who lives there. RoboShape, a new open-source tool from UTAustin SwarmLab, tries to draw a line inside the robot's perception: recognize the couch, but not that the couch is in a child's bedroom.

The tool is a small compression head bolted onto a frozen point cloud encoder. It trains the system to keep what helps object recognition and discard what would let a downstream model guess sensitive attributes such as room function or layout. The authors measure the trade three ways. The embeddings are 87.5% smaller than the underlying encoder's output. Object classification holds 98.7% of its original utility across three indoor LiDAR datasets. A probe trained to predict sensitive attributes loses 39.3% of its performance on those same datasets.

Those numbers come from the authors' own benchmark, against a specific probe class. The fix lives at the encoder, not the network. If the device runs inference locally or ships raw scans anyway, RoboShape does not help. The team releases the code on GitHub and frames the design as encoder-agnostic, so a privacy filter can travel with the robot rather than sit in a cloud policy.

The result is an open-source building block, not a privacy guarantee. For a reader choosing a home robot, the practical question is which parts of the room the device is allowed to remember. The encoder-side primitive is one answer; the rest depends on whether the device ships raw scans at all.