Beyond the Remembered World: Predictive 4D Belief for Persistent Navigation in Evolving Worlds

Beyond the Remembered World. Predictive 4D Belief for Persistent Navigation in Evolving Worlds. By Mingjian Gao and nine others, from Zhejiang University, the University of California, San Diego, and two other institutions. Published on the thirtieth of September. Its code repository is public and real: about thirty-five source files.

An embodied agent that keeps a persistent spatial memory can find its way around a familiar place over repeated visits. Memory, though, records where things were. Targets can move while nobody is watching, including while the agent is on its way, so a remembered location may be wrong by the time the agent gets there. Memory retrieval and state prediction have both advanced, but two things remain hard: accounting for the world continuing to change out of sight, and revising beliefs when visibility is limited. The authors frame this as Evolving-World Navigation. An agent infers where a target is from intermittent observations, predicts its state at the moment of inspection, and revises its beliefs using visual evidence.

EvolvingNav answers with a belief indexed by time. It builds that belief from timestamped 3D histories of objects, through a structured model of persistence and relocation. The belief keeps two possibilities apart: that a target has persisted at the place it was last seen, or that it has relocated somewhere else. It also holds back some probability for locations outside the set of known candidates, so the unexpected is never ruled out entirely.

An event-driven filter keeps the belief current. As time passes it propagates the belief forward, forecasts whether the target will occupy each candidate at the time it would be inspected, and folds in new RGB-D evidence. Not seeing the target counts as evidence too. A negative observation lowers the weight of a location hypothesis according to detection probabilities that are calibrated and conditioned on visibility, and evidence tracking stops the same observation from being used twice. Acting on all this is a frozen vision-language controller, used zero-shot, which reads the updated belief to choose actions and to replan.

To test the setting, the authors also introduce EvoWorld-Bench. It is grounded in traces of human activity and spans fifty-four scenes and eight hundred and three thousand six hundred and eighty tasks, with changes controlled both before and during navigation. In simulation and in experiments on a real robot, EvolvingNav improves navigation success and search efficiency over the baselines evaluated. Paired experiments find the clearest gains where the temporal patterns are learnable. Ablations show that keeping uncertainty alive, and taking visibility into account when weighing evidence, both earn their place.

A map of where things were is only half of what a returning agent needs. The other half is an expectation of where they have gone since, kept honest by what the agent sees and fails to see along the way. EvolvingNav treats memory as a forecast to be checked rather than a record to be trusted.