EvolvingNav Predicts Moving Targets in Changing Worlds

EvolvingNav uses a predictive 4D belief so robots can find targets that move while unobserved, beating baselines on a large benchmark and real robots.

EvolvingNav is a method for embodied navigation in worlds where targets can move while unobserved, including before or during a visit. It maintains a time-indexed belief from timestamped 3D object histories that distinguishes persistence and relocation; one report also says the belief covers locations outside the known set, while the other says an event-driven filter forecasts occupancy. Hypotheses are updated with visibility-conditioned RGB-D evidence without reusing the same observations, and a frozen zero-shot vision-language controller uses the belief to choose actions and replan. On EvoWorld-Bench, with 54 scenes and 803,680 tasks, and in real-robot tests, it improves success and search efficiency over baselines, especially when temporal movement patterns are learnable. The reports agree on the benchmark scale and the performance claim; they differ mainly in emphasizing unknown locations versus occupancy forecasting, and only the later report explicitly calls the benchmark changes controlled.

• EvolvingNav targets Evolving-World Navigation, where objects can move while unobserved before or during an agent’s visit.
• It builds a time-indexed belief from timestamped 3D object histories covering persistence and relocation.
• Sources differ on a further case: one includes locations outside the known set; the other describes an event-driven filter that forecasts occupancy.
• Visibility-conditioned RGB-D evidence revises hypotheses without reusing the same observations.
• A frozen zero-shot vision-language controller selects actions and replans from the belief.
• EvoWorld-Bench contains 54 scenes and 803,680 tasks; the later report says changes are controlled.
• It improves success and search efficiency versus baselines on that benchmark and in real-robot tests, with the largest gains when temporal relocation patterns are learnable.