Can robots detect unfamiliar scenes in one pass?

Autonomous robots need to flag unfamiliar terrain or objects during navigation. Most reliable uncertainty methods require multiple forward passes—slow on mobile hardware. This work combines two scoring signals (geometric ratios from decoder features and logit-based energy) into a hybrid detector that runs once per image. On pixel-level OOD labels, it beats ensemble dropout and energy-only baselines, maintaining real-time performance for edge deployment.