Write-safe flow mapping cuts ghost contamination by 42% in robotics

Mobile robots often rely on onboard sensors to infer flow structures like wind or water currents, but translating local estimates into a global map can go wrong. Ambiguous observations—similar flow patterns in different locations—combined with localization drift mean a robot might write a predicted velocity patch at the wrong position. Over time, these misregistered updates accumulate into persistent ghost structures that corrupt the map. Researchers from (likely Stanford) tackle this with a map-reference-aware conservative fusion framework. Their model predicts a local velocity patch and a learned write-safety score that continuously attenuates uncertain updates while still allowing initialization when no reliable map reference exists.

Testing in synthetic jet and crossflow environments, the method reduced average ghost contamination by 42% compared to ungated fusion. A zero-shot hardware replay using real pressure and optical-flow measurements from a thruster wake further showed robustness, cutting ghost contamination by 39% while retaining 81% map coverage. This work highlights that safe map writing—knowing when not to write—is as critical as accurate perception for building reliable flow maps. The approach could improve autonomous underwater vehicles, drones, and wind-field mapping systems where environmental flow data is noisy and localization degrades over time.