Object-goal navigation challenges an agent to find a target object in an unseen space and halt beside it. In practice, agents lean on object detectors, but those detectors can fire on ambiguous views. The moment to stop is not always clear from a single frame, and a premature stop can mean failure.
The paper introduces CLOSER, described as an evidence-aware active visual verification approach. Rather than trusting the detector’s first output, the agent is designed to assess whether the evidence it has collected justifies a stop, and to actively seek additional viewpoints when confidence is low.This shift from passive inference to active verification is meant to make navigation more reliable under weak or noisy perception.
The authors frame the method as a response to a core limitation in current ObjectNav systems: they often treat detection as final. CLOSER instead treats it as a hypothesis to be checked. The work is presented on arXiv as a new contribution to embodied navigation, though no experimental details or performance numbers are included in the available abstract.