Persistent spatial memory enables embodied agents to navigate familiar environments across repeated visits, but targets may move while unobserved, even during navigation, rendering remembered locations unreliable upon arrival. We formalize this as Evolving‑World Navigation: agents infer target locations from intermittent observations, predict their states at inspection time, and revise beliefs using visual evidence.
We introduce EvolvingNav, which builds a time‑indexed belief from timestamped 3D object histories via a structured persistence‑relocation model. The belief separates the probability of the target persisting at its last observed location from relocating to alternative candidates, while retaining probability mass for unknown locations.
An event‑driven filter propagates the current belief as time elapses, forecasts target occupancy at candidate inspection times, and incorporates new RGB‑D evidence. Negative observations down‑weight location hypotheses according to calibrated, visibility‑conditioned detection probabilities, and an evidence‑tracking mechanism prevents reuse of the same observation. A frozen, zero‑shot vision‑language controller then selects actions and replans based on the updated belief.
For evaluation we release EvoWorld‑Bench, a benchmark grounded in human activity traces, comprising 54 scenes and 803 680 tasks with controlled changes before and during navigation. Both simulation and real‑robot experiments show that EvolvingNav improves navigation success and search efficiency over all baselines. Paired experiments reveal the strongest gains under learnable temporal patterns, while ablations confirm the importance of preserving uncertainty and incorporating visibility‑aware evidence.
Review