AI

LifelongCrossNav Keeps a 3D Map Across Floors and Goals

Shar Hendrix 4 min read

Most object-goal bots either remember a flat map or climb stairs once for a single target. Real apartments ask for both: find the TV, then the bed, then the toilet—and maybe take the stairs between them.

LifelongCrossNav (arXiv:2608.07079, Aug 7, 2026) is a framework for sequential multi-object ObjectNav in unknown multi-floor indoor scenes. It keeps a shared sparse 3D semantic voxel memory across goals inside one episode, instead of rebuilding the world every time a new object name appears.

Project page: flageval-baai.github.io/LifelongCrossNavPage.

Cutaway multi-floor home with robot paths to three sequential goals
Sequential multi-object ObjectNav across floors: Start → Goal 1 → Goal 2 → Goal 3 with stair transitions. Source: arXiv:2608.07079.

What is new

  • Persistent 3D memory — geometry, traversability, and vision-language features accumulate while goals change.
  • Stairs as first-class structure — support-aware voxel types (Occupied / Traversable / Stair / Unsupported), stair-specific perception, direction-aware stair frontiers.
  • Unified policy — same-floor frontiers, stair frontiers, History POIs (re-query stored features with new goal text), Live POIs (YOLO / YOLO-World + MobileSAM).
  • HM3D-MFMON benchmark927 three-goal episodes from 36 multi-floor HM3D scenes, including 288 Cross-Floor-Required episodes that need at least one floor change.

“Lifelong” here means within-episode persistence across sequential goals, not lifelong learning across days.

LifelongCrossNav system framework diagram
Framework: support-aware 3D map + persistent VL memory + mode-aware 3D planning. Source: arXiv:2608.07079.

Results (vs planar OneMap baseline)

On all 927 episodes (paper Table 1; rows ordered as OneMap / w/o History-POI / full method in the HTML text):

  • Sequence SR rises from the planar baseline’s mid-teens toward roughly 29% for LifelongCrossNav variants
  • Progress rate (PR) approaches ~48%
  • History POIs mainly lift SPL / PPL (later goals reuse memory instead of re-exploring)

On the 288 Cross-Floor-Required episodes, the planar baseline completes 0% full sequences; LifelongCrossNav reaches nonzero sequence success (~8% SR class) with higher partial progress—planar semantic maps alone cannot finish tasks that require stairs.

Single-object HM3D ObjectNav remains competitive (59.7% SR, 28.0% SPL) without online LLM reasoning.

A Human’s Take

Stairs plus memory is how service robots stop looking lost between “find the remote” and “find the charging cable upstairs.” The Cross-Floor-Required zero for planar maps is the clearest receipt. I’m curious how History POIs behave when detection confuses beds and sofas—the authors already flag that as a real failure mode, which is the right kind of honesty.

Sources