cs.ROAug 24, 2026
SaveSuperMap: A Spatio-Temporal SLAM System for Visual-Language Navigation
Organizations: The Robotics Institute, Carnegie Mellon University
Abstract
Robotic navigation in human environments requires a spatio-temporal semantic representation that can rec- oncile open-vocabulary perception with long-term environmental changes. While foundation models provide strong zero-shot recognition, their predictions are intermittent and view-dependent, and naively integrating them into mapping pipelines leads to identity drift and stale semantics over time. We present SuperMap, a 4D spatio-temporal mapping framework for language-guided navigation that integrates high-frequency geometric SLAM with asynchronous open-vocabulary perception. Our core contribution is a consistency-driven mapping engine that combines 3D-aware instance association/re-activation with a principled existence-and-label confidence update to maintain stable object identities and prune outdated map content under occlusions and scene changes. SuperMap produces a queryable 4D scene-graph representation that interfaces naturally with Vision-Language Models by supporting compositional queries over object semantics, relations, We demonstrate SuperMap on benchmarks and real robots, including dynamic scenes with appearance/disappearance and relocation, and provide ablations and runtime analysis. We release the full system as open-source to provide the community with a deployable baseline for open-vocabulary spatio-temporal mapping. Project website: superodometry.com/supermap.
Figures & tables
| Capability | Semantic Mapping | Semantic SLAM | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| HOV-SG [ 30 ] | ConceptGraphs [ 5 ] | RayFronts [ 1 ] | OpenScene [ 16 ] | OpenMask3D [ 28 ] | SeeGround [ 11 ] | Kimera [ 24 ] | OVO-SLAM [ 15 ] | Khronos [ 25 ] | SuperMap (Ours) | |
| Real-Time | ✗ | ✗ | ✓ | ✗ | ✗ | ✗ | ✓ | ✓ | ✗ | ✓ |
| Open-Vocabulary | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✗ | ✓ | ✗ | ✓ |
| Instance-Level | ✓ | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ | ✓ | ✗ | ✓ |
| Short-Term | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | ✓ | ✓ |
| Long-Term | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | ✓ | ✓ |
TABLE I : Capability Checklist for Semantic Mapping and Semantic SLAM Methods. ✓indicates support , ✗indicates limited or no support
Fig. 2 : Overview . The system delivers accurate state estimation together with a rich, multi-level map representation that spans from geometric details to abstract semantics. It consists of three main modules. First, an online 3D reconstruction module built on a LiDAR SLAM backbone performs state estimation and generates a dense 3D model of the environment. Next, the estimated state and reconstructed geometry are used to identify short-term and long-term object changes, enabling instance-level object mapping. Finally, the object-level representation is organized into an abstract 4D scene graph that captures both spatial and temporal relationships among objects.
| Method | Mapping Approach | Without Background | With Background | ||||
|---|---|---|---|---|---|---|---|
| mIoU (%) | f-mIoU (%) | Acc (%) | mIoU (%) | f-mIoU (%) | Acc (%) | ||
| ConceptFusion [ 7 ] | point-feature | 21.76 | 26.71 | 34.10 | 18.57 | 23.06 | 28.77 |
| NACLIP-3D [ 6 ] | 22.32 | 24.32 | 33.46 | 22.32 | 24.32 | 33.46 | |
| Trident-3D [ 26 ] | 29.97 | 37.62 | 51.06 | 24.80 | 27.77 | 38.43 | |
| RayFronts [ 1 ] | 41.29 | 46.42 | 56.76 | 32.29 | 39.04 | 49.15 | |
| ConceptGraphs [ 5 ] | object-level | 21.62 | 24.32 | 31.05 | 20.83 | 23.61 | 35.80 |
TABLE II : Class-level Segmentation Benchmarking on ScanNet. Best results are highlighted as first and second .
| Method | Chair | Window | Refrigerator | Sofa | Door | |||||
|---|---|---|---|---|---|---|---|---|---|---|
| mAP 50 | mAP 25 | mAP 50 | mAP 25 | mAP 50 | mAP 25 | mAP 50 | mAP 25 | mAP 50 | mAP 25 | |
| HOV-SG [ 30 ] | 4.58 | 4.73 | 0.00 | 0.00 | 0.00 | 0.00 | 30.00 | 31.25 | 9.70 | 10.40 |
| ConceptGraphs [ 5 ] | 0.00 | 2.33 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 |
| SuperMap (Ours) | 63.76 | 74.72 | 42.20 | 67.92 | 62.50 | 62.50 | 33.35 | 83.35 | 10.00 | 25.00 |
TABLE III : Instance-level Segmentation Benchmark on ScanNet. Best results are highlighted as first and second .
Fig. 3 : Spatio-temporal Consistency. For each example (left to right), we show the 2D detections, the dense 3D reconstruction, and the object-centric map rendered in Rerun before and after the scene change. Top: three appearance events (bucket, cart, safety sign), where newly introduced objects receive new IDs while nearby objects retain their original IDs, demonstrating long-horizon re-identification and stable instance association in 3D. Bottom: three disappearance events (chair, plant, trash can), where the same instance ID is preserved across time despite the object leaving the field of view.
| Method | Task | Appeared | Disappeared | ||||
|---|---|---|---|---|---|---|---|
| Buc. | Cart | Sign | Plant | Trash | Chair | ||
| Khronos [ 25 ] | Detect. | – | – | – | – | – | – |
| DualMap [ 8 ] | 0.000 | 0.000 | 0.000 | 0.000 | 0.310 | 0.000 | |
| SuperMap (ours) | 1.000 | 0.262 | 0.583 | 0.755 | 0.434 | 1.000 | |
| Khronos [ 25 ] | Change | – | – | – | – | – | – |
| DualMap [ 8 ] | 0.553 | 0.527 | 0.507 | 0.561 | 0.642 | 0.449 | |
TABLE IV : Spatial–Temporal Change Detection Performance.
| Configuration | Precision | Recall | F1 |
|---|---|---|---|
| W/o 2D Tracker | 0.7787 | 0.4595 | 0.5780 |
| W/o Semantic Fusion | 0.7929 | 0.3870 | 0.5201 |
| W/o Geometric Consistency Update | 0.8189 | 0.4448 | 0.5764 |
| All (proposed) | 0.8677 | 0.4955 | 0.6308 |
TABLE V : Performance Comparison of Ablation Study
Fig. 4 : Comparison of Reasoning Results using Video Input and Scene-graph Input. Compared to raw video, the scene-graph representation yields more reliable spatial reasoning, clearer temporal inference, and more accurate instance-level object retrieval.
Fig. 5 : Online Visual–Language Navigation using Scene Graphs. Our system enables robots to understand and reason about the spatial relationships in language instructions for VLN. See the supplementary video for more details.
| Method | Chair | Window | Refrigerator | Sofa | Door | Avg. Mem (MB) | Peak Mem (MB) | TPF (s) | |||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| mAP 50 | mAP 25 | mAP 50 | mAP 25 | mAP 50 | mAP 25 | mAP 50 | mAP 25 | mAP 50 | mAP 25 | ||||
| HOV-SG [ 30 ] | 4.58 | 4.73 | 0.00 | 0.00 | 0.00 | 0.00 | 30.00 | 31.25 | 9.70 | 10.40 | 8755.86 | 10226.66 | 8.623 |
| ConceptGraphs [ 5 ] | 0.00 | 2.33 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 859.86 | 886.18 | 0.092 |
| SuperMap (Ours) | 63.76 | 74.72 | 42.20 | 67.92 | 62.50 | 62.50 | 33.35 | 83.35 | 10.00 | 25.00 | 2030.09 | 2184.86 | 0.3604 |
TABLE I : 3D instance segmentation on ScanNet with memory and runtime statistics. TPF denotes mapping time per frame. Best results are highlighted as first and second .
Fig. 1 : Ablation on tracking under large viewpoint changes. Our 2D–3D tracker (top) preserves consistent object tracks and yields a coherent map across significant viewpoint changes, whereas the 2D-only tracker (bottom) suffers from track fragmentation, resulting in incorrect object identities. Notably, the 2D-only tracker (bottom) exhibits clear ID switches between t = 0 s and t = 2.5 s.
Fig. 2 : Ablation on tracking dynamic objects. The first row shows our 2D–3D tracking, while the second row shows 2D-only tracking. The 2D-only tracker frequently loses the target, resulting in the same person being represented as multiple objects in the map. In contrast, our 2D–3D tracker maintains a consistent identity over time.
Fig. 3 : Ablation of the consistency check. Comparison of mapping results without consistency checks, with geometric-only checks, and with joint geometric+semantic checks (ours). Without consistency checks, stale points from dynamic objects persist in the map. Geometric checks remove some unsupported points but may leave projection artifacts. The joint geometric and semantic consistency check produces a cleaner and more up-to-date map.
Fig. 4 : Qualitative comparison of instance-level segmentation. Background points are removed for clarity. SuperMap maintains more consistent instance identities and shows fewer confusions between objects and background compared to prior methods.
Fig. 5 : Newly appeared object (yellow bucket). When revisiting the environment, SuperMap detects the newly introduced bucket and incorporates it into the object-level map.
Fig. 6 : Newly appeared object (cart). The system recognizes the cart as a new instance and updates the map accordingly.
Fig. 7 : Newly appeared object (safety sign). SuperMap adds the newly observed sign to the spatio-temporal map.
Fig. 8 : Disappeared object (plant on table). The system removes the plant from the map after it is no longer observed.
Fig. 9 : Disappeared object (blue trash can). SuperMap updates the map to reflect the removal of the trash can.
Fig. 10 : Disappeared object (chair). The system correctly removes the chair from the object-level map once it is absent.