Beyond the Remembered World: Predictive 4D Belief for Persistent Navigation in Evolving Worlds
EvolvingNav keeps searching agents from trusting stale object locations when the world changes out of sight.
The paper frames “Evolving-World Navigation” as a setting where targets can move before or during navigation, so agents must update beliefs from intermittent RGB-D evidence. Its method builds a time-indexed belief over object histories, separating likely persistence from relocation and keeping uncertainty outside known candidates. Negative observations reduce location hypotheses based on visibility-conditioned detection probabilities. The authors also introduce EvoWorld-Bench, with 54 scenes and 803,680 tasks, and report better navigation success and search efficiency than evaluated baselines. HF Daily Papers' note
The paper frames “Evolving-World Navigation” as a setting where targets can move before or during navigation, so agents must update beliefs from intermittent RGB-D evidence. Its method builds a time-indexed belief over object histories, separating likely persistence from relocation and keeping uncertainty outside known candidates. Negative observations reduce location hypotheses based on visibility-conditioned detection probabilities. The authors also introduce EvoWorld-Bench, with 54 scenes and 803,680 tasks, and report better navigation success and search efficiency than evaluated baselines. HF Daily Papers' note
score 4