SuperMap: A Spatio-Temporal SLAM System for Visual-Language Navigation

3 min read

SuperMap keeps an open-vocabulary, instance-level map of objects that stays valid as things are moved, removed or added, and exposes it as a “4D” scene graph a vision-language model can query. Geometry and pose come from SuperOdometry (lidar, camera and IMU). GroundingDINO and SAM2 give 2D detections at about 1 Hz. Each 2D track runs a Kalman filter whose prior is the object’s 3D centroid projected with the current pose, which keeps identities through occlusion and fast turns. Each map point keeps a log-odds occupancy that a depth residual updates: a point now in front of the measured surface counts as gone. Labels are fused as a categorical belief using the detector’s confusion matrix. The graph has spatial edges from class-dependent geometric rules (on, beside, under) and temporal edges along each object’s track. It is serialised to text for Gemini 2.0 Flash, which returns instance IDs that become navigation waypoints.

Findings

  • ScanNet class-level segmentation, background excluded: accuracy 55.48%, 1.32 points below RayFronts, and ahead of ConceptGraphs (31.05%) and HOV-SG (35.17%). Its mIoU is lower (27.42 against 41.29 for RayFronts), which the text does not dwell on.
  • ScanNet instance segmentation: chair mAP50 63.76 against 4.58 for HOV-SG and 0.00 for ConceptGraphs. ConceptGraphs scores zero on almost every class, which suggests the comparison protocol suits SuperMap’s tracking more than the baselines.
  • A 10-minute real run over a 30 m by 20 m indoor area, with three objects removed and three added: change-detection recall from 0.622 (cart) to 1.000 (bucket, chair). Detection recall for the cart is only 0.262. DualMap’s detection is near zero, and Khronos gave no usable output in their runs.
  • Ablation on 41 annotated objects: full system F1 0.6308, against 0.5201 without semantic fusion. Recall stays below 0.5 even for the full system.
  • Rates on an i9-14900H with an RTX 4090 laptop GPU: pose 10 Hz, 2D segmentation 1 Hz, 3D mapping 3 Hz, graph updates 5 Hz.
  • The language-navigation demos are qualitative: picking the right whiteboard out of four identical ones, and finding the painting next to the fridge.

Limits

  • Fast-moving objects are tracked poorly (authors).
  • Detection runs from a predefined list of object prompts, so “open vocabulary” means open within that list (authors).
  • The change benchmark is six objects in one sequence, and the graph-against-video comparison for VLM reasoning is two worked prompts.
  • The disappearance test needs a depth return from behind where the object stood, the same reference-surface assumption Khronos states and this paper leaves implicit.
  • One robot with a Livox Mid-360 and a panoramic camera, aligned by software timestamps. Nothing on multiple robots, sharing or bad links.
  • It says Khronos “struggles to operate in real time”, while Khronos reports 22.2 FPS for its active window with segmentation precomputed. The claims probably differ on whether segmentation is counted.

Relevance

SuperMap and Khronos agree that a long-lived map needs stable object identity plus an explicit test for absence. SuperMap adds open vocabulary and a language interface on top; see long-term autonomy and open-vocabulary mapping. How its existence beliefs would behave when the evidence comes from several robots with different sensors is an open question.