Robots use active perception to disambiguate targets without asking users

Natural language gives robots a flexible interface, but target ambiguity often stems from missing physical evidence—occlusion, restricted viewpoints, or unreadable labels—not just unclear user intent. Existing interactive disambiguation methods typically ask the user for more information, which fails when the user doesn't know what the robot can't see. To address this, Yiwei Liu and Luwei Yang from arXiv paper 2608.13605 introduce an active-perception framework that treats movement as the primary tool for information acquisition.

At its core, the framework uses a vision-language model to fuse accumulated visual evidence with interaction data, then makes a high-level decision: continue observing, request clarification, or complete target selection. This lets the robot directly recover discriminative evidence (like object names, labels, or semantic attributes) by repositioning itself, making any necessary user questions more targeted and effective. The authors validated the approach in real-robot experiments, showing that combining physical observation with user intent clarification in a unified process outperforms methods that rely solely on dialogue. This moves beyond passive help-seeking toward truly embodied disambiguation.