Robots that remember a room are only half the story - what happens when something moves while you're not looking?
Researchers built EvolvingNav, a navigation system that tracks timestamped histories of objects and forecasts where they are likely to be by the time a robot actually arrives. The system splits its guess into two cases: the object stayed where it was last seen, or it relocated somewhere else, and it keeps some probability mass on "somewhere not yet considered" rather than forcing a single answer. An event-driven filter updates that belief as time passes, folds in new RGB-D camera evidence, and downweights locations using detection probabilities calibrated to how visible an object actually is from a given vantage point. A separate, frozen vision-language model then picks actions and replans based on the updated belief. The team tested it on EvoWorld-Bench, a new benchmark built from human activity traces spanning 54 scenes and 803,680 tasks, plus real-robot trials, and reported better navigation success and search efficiency than the baseline systems they compared against.
Most navigation research still treats a mapped room as basically frozen between visits. That is a bad assumption for home or office robots, where someone moves a mug, a delivery lands at a different door, or a cart gets wheeled to another aisle. Treating "I have not seen it recently" as genuine uncertainty, rather than stale fact, is the more interesting design choice here than the engineering details.
The catch: the benchmark is synthetic, built from scripted human-activity traces, not the real chaos of a household over weeks. Whether this scales to environments with hundreds of objects moving at once, without the compute cost of maintaining a time-indexed belief for each one, is the question the paper leaves open.