Continuous aerial search requires building target hypotheses from sparse, viewpoint-dependent observations. A known failure mode is false candidate persistence (FCP): when later evidence contradicts a stored hypothesis, the candidate can remain available to the planner and keep triggering approach or verification actions. The authors of a new arXiv paper identify this trajectory-level failure and propose SkyAgent, a system equipped with Revisable Target Memory (RTM).
RTM associates evidence with persistent candidates, discounts correlated views, and revises each candidate through Tentative, Confirmed, and Rejected states. The key interface binds every generated waypoint to its source candidate, so rejection invalidates stale goals while conservative reactivation protects temporarily occluded targets. This state-to-action correction is designed to prevent wasted travel and behavior-level false positives.
Across four splits from the CityAVOS, UAV-ON, and OpenUAV benchmarks, SkyAgent with RTM outperforms all competitors, improving success rate by 2.90–6.31 points while substantially reducing Behavior-FCP and post-contradiction travel, with a low false-rejection rate. In 35 matched real-world DJI scenarios per method, RTM consistently improved task success and behavioral reliability, validating the approach under physical deployment conditions.