How could persistent 3D semantic memory for navigation change multi-floor object navigation over the next two years?

What changed: from forgetting each floor to remembering the whole building

Previously, object navigation assumed a single floor and often rebuilt a map from scratch for each new object query. That meant a robot searching for a cup, then a book, would re-explore the same rooms, wasting time and energy. The new approach is persistent 3D semantic memory: the robot maintains a shared 3D voxel map that accumulates geometric structure, traversability, and vision-language features as it moves [1]. When a new object goal arrives, the robot queries this memory instead of starting over, so it can go directly to a previously seen location or know where it still needs to look.

Evidence from LifelongCrossNav shows this works in practice: on a new benchmark for sequential multi-floor multi-object navigation (HM3D-MFMON), the persistent 3D memory system consistently outperformed a planar persistent-map baseline [1]. The key is that the memory is 3D and shared across subtasks, so the robot can reuse information about stairs, rooms, and object locations across the whole episode. This is a fundamental shift from 'explore and forget' to 'explore and remember'.

The hard part: stairs, elevators, and vertical reasoning

Multi-floor navigation isn't just about memory—it's about knowing how to move between floors. Older systems either ignored stairs or required a pre-built map. New frameworks explicitly model stairways and floor transitions. LifelongCrossNav combines support-aware 3D traversability mapping, stair-specific perception, and direction-aware stair traversal [1]. Similarly, the ASCENT framework builds a hierarchical representation with stair-aware obstacle mapping and cross-floor topology, enabling online, floor-aware navigation without pre-built maps [5]. These systems treat stairs as first-class navigation elements, not obstacles.

The payoff is measurable: ASCENT outperformed state-of-the-art zero-shot approaches on HM3D and MP3D benchmarks [5], and the Multi-Floor Navigation Policy (MFNP) achieved higher success rates and better exploration efficiency than existing zero-shot methods [3]. Real-world tests on a quadruped robot confirmed that these policies can find objects in completely unseen multi-floor buildings [3]. So over the next two years, expect robots to handle stairs and floor changes as routine, not exceptional.

What this means for real robots in the next two years

The practical impact is that service and delivery robots will become more useful in real buildings—offices, hospitals, homes—where multi-floor layouts are the norm. Instead of requiring a pre-built map or retraining for each new object category, these systems work online and zero-shot, meaning they can handle new objects and new buildings without prior data [5]. The memory also supports multi-modal queries (e.g., 'find the red mug' or 'go to the kitchen') by storing vision-language features [1][4].

However, there are caveats. Most results are from simulated benchmarks (HM3D, MP3D), and real-world tests are limited to a few quadruped robots [3][5]. The BUG-FIRE study, focused on firefighter navigation, found that a 30-meter vision radius reduced path length by 28.2% compared to 0.5-meter visibility, highlighting that perception range still matters [2]. Also, floor-plan priors can help, as FloVerse showed that using floor plans improves navigation across all goal types [6]. So while the next two years will bring more capable multi-floor navigation, expect continued improvements in perception and robustness before widespread deployment.

About These Sources

This answer is built on 6 studies (4 peer-reviewed, 2 preprints) — published from 2025 to 2026, 6 from 2024 or later, 2 in Q1–Q2 journals — selected as the most relevant from 7 studies that passed quality screening, drawn from 23 papers retrieved from a database of over 500 million.

Sources used in this answer

Related research questions