pith. sign in

arxiv: 2606.31471 · v1 · pith:OOUM44PBnew · submitted 2026-06-30 · 💻 cs.CV

Think While You Map: Asynchronous Vision-Language Agents for Incremental 3D Scene Graphs

Deniz Bickici , Michael Pabst , Shohei Mori , Dieter Schmalstieg This is my paper

Pith reviewed 2026-07-01 06:13 UTC · model grok-4.3

classification 💻 cs.CV
keywords 3D scene graphsopen-vocabulary segmentationvision-language modelsincremental mappingvisual groundingasynchronous agents
0
0 comments X

The pith

Asynchronous vision-language agents enrich 3D scene graphs incrementally while mapping proceeds.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that open-vocabulary 3D scene graph construction does not need to wait until mapping finishes before adding language-based semantics. Instead, a lightweight probabilistic voxel mapping runs at the same time as heavier vision-language model agents that refine the graph in the background. This makes the graph usable for queries right away and leads to better results on semantic segmentation and visual grounding tasks compared to previous methods.

Core claim

An asynchronous framework separates lightweight online mapping from heavyweight semantic refinement by vision-language models, allowing the scene graph to be queryable during exploration while progressively gaining semantic richness through semantic loop closure and attribute attachment.

What carries the argument

probabilistic voxel-based backbone that maintains stable object identities incrementally, combined with background VLM agents and a multi-target frame scheduler

If this is right

  • The resulting scene graph is queryable during exploration.
  • It resolves duplicate object tracks through semantic loop closure.
  • Fine-grained visual attributes and spatial relations are derived between objects.
  • It matches or outperforms existing methods on semantic segmentation on ScanNet and Replica.
  • It surpasses prior state-of-the-art on visual grounding benchmarks by 15.3 to 18.8 A@0.25.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This approach could support robots that need to understand and query their environment in real time without waiting for post-processing.
  • Amortizing VLM costs over informative frames might generalize to other expensive perception tasks.
  • Semantic loop closure could be tested in longer, more complex environments to check for drift accumulation.

Load-bearing premise

The probabilistic voxel-based backbone maintains stable object identities incrementally without drift that would require later correction by the VLM agents.

What would settle it

Observing significant object identity drift in extended mapping sequences on datasets like ScanNet without corresponding VLM corrections would falsify the stability claim.

Figures

Figures reproduced from arXiv: 2606.31471 by Deniz Bickici, Dieter Schmalstieg, Michael Pabst, Shohei Mori.

Figure 1
Figure 1. Figure 1: ThinkGraphs decouples lightweight online mapping from heavyweight VLM reasoning, keeping the scene graph queryable during exploration. Asynchronous back￾ground agents refine the graph without blocking the mapping loop: a Critic Agent detects and merges fragmented object tracks, and a Description Agent injects fine￾grained visual attributes, enabling complex grounding queries (e.g., “the stainless steel ref… view at source ↗
Figure 2
Figure 2. Figure 2: Method Overview. (i) The frontend extracts grounded instances from each RGB-D frame (Sec. 3.1). (ii) The backend associates them into persistent 3D tracks with probabilistic voxel scoring and derives spatial edges deterministically from 3D geometry (Sec. 3.2). (iii) Two asynchronous VLM agents, a Critic Agent for semantic loop closure and a Description Agent for attribute enrichment, progressively refine t… view at source ↗
Figure 3
Figure 3. Figure 3: VLM Scheduling. Multi-target frame scheduling for the Description Agent (left) and pair scheduling for the Critic Agent (right), both using Set-of-Mark overlays. detection confidence q is associated to Ti , the histogram is updated as \mathcal {H}_{i}(\ell ) \;\leftarrow \; \mathcal {H}_{i}(\ell ) + q, \label {eq:label_vote} (3) and the consensus label is the highest-scoring label: \ell _i^\star = \underse… view at source ↗
Figure 4
Figure 4. Figure 4: Incremental Segmentation. On Room0 (Replica), metrics converge steadily as frames are processed. Tab. 1 compares our method with prior open-vocabulary 3D methods on Replica and ScanNet. On Replica, our incremental method outperforms all prior methods on every metric (0.58 mAcc, 0.37 mIoU, 0.61 f-mIoU), with the largest gain on f-mIoU (+0.13 over BBQ-CLIP). On ScanNet, we obtain 0.75 mAcc, 0.44 mIoU, and 0.… view at source ↗
Figure 5
Figure 5. Figure 5: Effect of the Critic Agent. Without the Critic (left), association drift frag￾ments objects into duplicate object tracks (red circles). Semantic loop closure (right) merges them into consistent identities. Points are colored by instance. architecture from the frontend VLM (a lighter RAM++ tagger) is provided in the supplementary, showing the gains are not merely due to a stronger tagger. Runtime and cost. … view at source ↗
Figure 6
Figure 6. Figure 6: Critic Agent Decisions. Examples on ScanNet; green borders mark correct decisions, red borders errors. a correct rejection. Two visually similar boxes stacked vertically are kept as distinct instances due to their low containment (7.1%) and vertical separation (0.23 m). Conversely, this is a case where IoU- or containment-based merging would erroneously fuse two distinct same-class instances; the Critic ke… view at source ↗
Figure 7
Figure 7. Figure 7: Description Agent Call. A single multi-target call on ScanNet scene0389_00: three SoM-annotated frames jointly cover six objects, refining coarse labels into fine-grained descriptions in one VLM call. disambiguation of same-class instances during grounding. In one case (track_2), the agent incorrectly overrides the correct consensus label clothes hanger with wall-mounted reading lamp. Since the backend con… view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative Segmentation Results. Open-vocabulary semantic segmenta￾tion on three Replica scenes: ground truth (top), our prediction (bottom). Asynchronous agents. Both VLM agents run in dedicated background threads, decoupled from per-keyframe processing. On average, the Description Agent takes 22.7 s to complete, with a mean lag of 28.9 keyframes (∼289 original frames at stride 10). The Critic Agent take… view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative Grounding Results. On Nr3D, Sr3D+, and ScanRefer. Green denotes the ground-truth and red the predicted bounding box. All shown examples satisfy A@0.25. Why Qwen3-VL as the proposal model? Grounded-SAM requires a set of textual object candidates for each frame. Since label stability has a significant downstream impact, the choice of proposal model is critical to avoid the cascade of noisy or ove… view at source ↗
Figure 10
Figure 10. Figure 10: Grounded-SAM vs. SAM2. Grounded-SAM (left) uses text-conditioned queries and yields fewer, object-aligned masks; SAM2 (right) produces class-agnostic masks that fragment objects. H Prompts We provide the exact prompts used for the frontend tagger, asynchronous VLM agents, and grounding evaluator. Frontend Tagging Prompt –- User –- List the distinct objects you can see in this scene. Respond with a comma-s… view at source ↗
read the original abstract

Open-vocabulary 3D scene graph methods typically operate in two stages: first reconstruct, then enrich with vision-language models, leaving the graph unqueryable during exploration. We argue that this sequential coupling is unnecessary and propose an asynchronous architecture in which lightweight online mapping runs concurrently with heavyweight semantic refinement. A probabilistic voxel-based backbone maintains stable object identities incrementally, while background VLM agents progressively enrich the graph. This framework resolves duplicate object tracks through semantic loop closure, attaches fine-grained visual attributes and derives spatial relations between objects. A multi-target frame scheduler amortizes VLM cost by selecting a small set of informative frames that jointly cover multiple targets. The resulting scene graph is queryable during exploration and grows in semantic richness over time. Our method matches or outperforms existing open-vocabulary 3D scene graph methods on semantic segmentation (ScanNet, Replica) and surpasses the prior state-of-the-art across three visual grounding benchmarks (Sr3D+, Nr3D, ScanRefer) by 15.3 to 18.8 A@0.25. Project page: https://denizbickici.github.io/thinkgraphs/

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes an asynchronous architecture for open-vocabulary 3D scene graph construction in which a lightweight probabilistic voxel-based backbone performs incremental online mapping to maintain object identities, while background VLM agents enrich the graph with semantics, attributes, and relations. Semantic loop closure resolves duplicates, and a multi-target frame scheduler selects informative frames to amortize VLM cost. The resulting graph is claimed to be queryable during exploration. Experiments report that the method matches or exceeds prior open-vocabulary scene graph methods on semantic segmentation (ScanNet, Replica) and surpasses prior SOTA on three visual grounding benchmarks (Sr3D+, Nr3D, ScanRefer) by 15.3–18.8 A@0.25.

Significance. If the asynchronous decoupling and identity stability hold, the work offers a practical route to queryable scene graphs during exploration rather than post-hoc reconstruction, which could benefit robotics and AR systems that require online semantic queries. The reported benchmark gains on standard datasets provide initial evidence of utility, though the absence of ablations and identity-stability metrics limits assessment of whether gains derive from the architecture or implementation choices.

major comments (3)
  1. [Method (backbone description) and Experiments] The central claim that the probabilistic voxel backbone 'maintains stable object identities incrementally' (abstract and method overview) lacks any reported quantitative metric of identity persistence, such as ID-switch rate, voxel-to-object assignment consistency across frames, or drift statistics over full trajectories. Without this, it is unclear whether the reported grounding gains require the online asynchronous property or could be achieved with offline post-hoc cleanup.
  2. [Experiments] No ablation studies isolate the contribution of semantic loop closure versus the VLM agents or the frame scheduler (Experiments section). This makes it difficult to determine whether the 15.3–18.8 A@0.25 gains on grounding benchmarks are attributable to the asynchronous design or to other factors such as VLM choice or post-processing.
  3. [Experiments and Discussion] The manuscript provides no error analysis or failure-case breakdown on the grounding benchmarks (e.g., cases where identity drift would force retroactive corrections). This is load-bearing for the claim that the graph remains reliably queryable during exploration.
minor comments (2)
  1. [Method] Notation for the probabilistic voxel representation and the multi-target scheduler could be clarified with a single diagram or pseudocode listing the key variables.
  2. [Abstract and Introduction] The abstract and introduction would benefit from an explicit statement of the computational budget (e.g., VLM calls per second or frame selection rate) to contextualize the amortization claim.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. The comments highlight important gaps in quantitative validation of the backbone's identity stability, component contributions, and error analysis. We will revise the manuscript to address these points directly while preserving the core asynchronous architecture and reported results.

read point-by-point responses

  1. Referee: [Method (backbone description) and Experiments] The central claim that the probabilistic voxel backbone 'maintains stable object identities incrementally' (abstract and method overview) lacks any reported quantitative metric of identity persistence, such as ID-switch rate, voxel-to-object assignment consistency across frames, or drift statistics over full trajectories. Without this, it is unclear whether the reported grounding gains require the online asynchronous property or could be achieved with offline post-hoc cleanup.

    Authors: We agree that explicit identity-persistence metrics would strengthen the presentation. The probabilistic voxel backbone maintains object identities through per-voxel Bayesian occupancy and label updates that favor temporal consistency, with semantic loop closure explicitly resolving duplicate tracks. The grounding benchmark gains (which penalize identity errors) provide indirect support for stability under the online regime. In revision we will add ID-switch rate, voxel-to-object assignment consistency, and trajectory drift statistics computed on the ScanNet and Replica sequences to quantify this property and clarify the contribution of the asynchronous design. revision: yes

  2. Referee: [Experiments] No ablation studies isolate the contribution of semantic loop closure versus the VLM agents or the frame scheduler (Experiments section). This makes it difficult to determine whether the 15.3–18.8 A@0.25 gains on grounding benchmarks are attributable to the asynchronous design or to other factors such as VLM choice or post-processing.

    Authors: We will add the requested ablations in the revised Experiments section. These will include controlled variants that disable semantic loop closure, the background VLM agents, and the multi-target frame scheduler individually while keeping all other components fixed, allowing direct measurement of each module's contribution to the reported grounding improvements. revision: yes

  3. Referee: [Experiments and Discussion] The manuscript provides no error analysis or failure-case breakdown on the grounding benchmarks (e.g., cases where identity drift would force retroactive corrections). This is load-bearing for the claim that the graph remains reliably queryable during exploration.

    Authors: We will include a new error-analysis subsection in the revised manuscript. It will present a breakdown of failure cases on Sr3D+, Nr3D, and ScanRefer, with particular attention to instances of potential identity drift and how semantic loop closure and the asynchronous update schedule mitigate retroactive corrections, thereby supporting the claim of reliable online queryability. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical systems paper with independent benchmark claims

full rationale

The paper describes an asynchronous architecture using a probabilistic voxel backbone for incremental mapping and background VLM agents for semantic enrichment, with results on semantic segmentation (ScanNet, Replica) and visual grounding (Sr3D+, Nr3D, ScanRefer). No equations, fitted parameters, self-definitional relations, or load-bearing self-citations appear in the provided text that would reduce any claimed performance or property to a construction from its own inputs. The contribution is presented as an empirical systems integration whose validity rests on external benchmark comparisons rather than internal redefinition or renaming of known results. The derivation chain is therefore self-contained against external evaluation.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; the voxel backbone and VLM agents are treated as standard components.

pith-pipeline@v0.9.1-grok · 5737 in / 1119 out tokens · 27610 ms · 2026-07-01T06:13:54.390856+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

56 extracted references · 44 canonical work pages · 6 internal anchors

  1. [1]

    In: Computer Vision – ECCV 2020, vol

    Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: ReferIt3D: Neural Listeners for Fine-Grained 3D Object Identification in Real-World Scenes. In: Computer Vision – ECCV 2020, vol. 12346, pp. 422–440. Springer International Publishing, Cham (2020).https://doi.org/10.1007/978-3-030-58452-8_25

    work page doi:10.1007/978-3-030-58452-8_25 2020

  2. [2]

    Qwen3-VL Technical Report

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., Ge, W., Guo, Z., Huang, Q., Huang, J., Huang, F., Hui, B., Jiang, S., Li, Z., Li, M., Li, M., Li, K., Lin, Z., Lin, J., Liu, X., Liu, J., Liu, C., Liu, Y., Liu, D., Liu, S., Lu, D., Luo, R., Lv, C., Men, R., Meng, L., Ren, X., Ren, X., Song, S., Sun, Y., Tang, ...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2511.21631 2025

  3. [3]

    IEEE Robotics and Automation Letters8(8), 4927–4934 (2023).https://doi.org/10

    Bavle,H.,Sanchez-Lopez,J.L.,Shaheer,M.,Civera,J.,Voos,H.:S-Graphs+:Real- Time Localization and Mapping Leveraging Hierarchical Representations. IEEE Robotics and Automation Letters8(8), 4927–4934 (2023).https://doi.org/10. 1109/LRA.2023.3290512 16 D. Bickici et al

    work page arXiv 2023

  4. [4]

    In: IEEE International Conference on Robotics and Automation, ICRA 2023, London, UK, May 29 - June 2, 2023

    Chen, B., Xia, F., Ichter, B., Rao, K., Gopalakrishnan, K., Ryoo, M., Stone, A., Kappler, D.: Open-vocabulary Queryable Scene Representations for Real World Planning. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 11509–11522. IEEE, London, United Kingdom (2023).https://doi. org/10.1109/ICRA48891.2023.10161534

    work page doi:10.1109/icra48891.2023.10161534 2023

  5. [5]

    In: Computer Vision – ECCV 2020, vol

    Chen, D.Z., Chang, A.X., Nießner, M.: ScanRefer: 3D Object Localization in RGB- D Scans Using Natural Language. In: Computer Vision – ECCV 2020, vol. 12365, pp. 202–221. Springer International Publishing, Cham (2020).https://doi.org/ 10.1007/978-3-030-58565-5_13

    work page doi:10.1007/978-3-030-58565-5_13 2020

  6. [6]

    In: 2017 IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR)

    Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scan- Net: Richly-Annotated 3D Reconstructions of Indoor Scenes. In: 2017 IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR). pp. 2432–2443. IEEE, Honolulu, HI (2017).https://doi.org/10.1109/CVPR.2017.261

    work page doi:10.1109/cvpr.2017.261 2017

  7. [7]

    In: 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

    Deng, Y., Yao, B., Tang, Y., Zhou, T., Yang, Y., Yue, Y.: Openvox: Real- time instance-level open-vocabulary probabilistic voxel representation. In: 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1305–1311 (2025).https://doi.org/10.1109/IROS60139.2025.11246455

    work page doi:10.1109/iros60139.2025.11246455 2025

  8. [8]

    In: Proceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD)

    Ester, M., Kriegel, H.P., Sander, J., Xu, X.: A density-based algorithm for discov- ering clusters in large spatial databases with noise. In: Proceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD). pp. 226–231. AAAI Press, Portland, Oregon (1996)

    1996

  9. [9]

    Image Vision Comput.149(C) (2024).https: //doi.org/10.1016/j.imavis.2024.105171

    Fang, Y., Sun, Q., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva-02: A visual representation for neon genesis. Image Vision Comput.149(C) (2024).https: //doi.org/10.1016/j.imavis.2024.105171

    work page doi:10.1016/j.imavis.2024.105171 2024

  10. [10]

    IEEE Transactions on Circuits and Systems for Video Technology35(8), 7667–7682 (2025).https://doi.org/10

    Feng, M., Yan, C., Wu, Z., Dong, W., Wang, Y., Mian, A.: History-enhanced 3d scene graph reasoning from rgb-d sequences. IEEE Transactions on Circuits and Systems for Video Technology35(8), 7667–7682 (2025).https://doi.org/10. 1109/TCSVT.2025.3548308

    work page arXiv 2025

  11. [11]

    Gu, Q., Kuwajerwala, A., Morin, S., Jatavallabhula, K.M., Sen, B., Agarwal, A., Rivera, C., Paul, W., Ellis, K., Chellappa, R., Gan, C., De Melo, C.M., Tenenbaum, J.B., Torralba, A., Shkurti, F., Paull, L.: ConceptGraphs: Open-Vocabulary 3D SceneGraphsforPerceptionandPlanning.In:2024IEEEInternationalConference on Robotics and Automation (ICRA). pp. 5021–5...

    work page doi:10.1109/icra57147.2024.10610243 2024

  12. [12]

    In: IEEE International Conference on Robotics and Automation, ICRA 2023, London, UK, May 29 - June 2, 2023

    Huang, C., Mees, O., Zeng, A., Burgard, W.: Visual Language Maps for Robot Navigation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 10608–10615 (2023).https://doi.org/10.1109/ICRA48891.2023. 10160969

    work page doi:10.1109/icra48891.2023 2023

  13. [13]

    In: Pro- ceedings of the 33rd ACM International Conference on Multimedia (ACM MM)

    Huang, X., Huang, Y.J., Zhang, Y., Tian, W., Feng, R., Zhang, Y., Xie, Y., Li, Y., Zhang, L.: Open-set image tagging with multi-grained text supervision. In: Pro- ceedings of the 33rd ACM International Conference on Multimedia (ACM MM). p. 4117–4126. Association for Computing Machinery, New York, NY, USA (2025). https://doi.org/10.1145/3746027.3755316

    work page doi:10.1145/3746027.3755316 2025

  14. [14]

    In: Robotics: Science and Systems XVIII

    Hughes, N., Chang, Y., Carlone, L.: Hydra: A Real-time Spatial Perception System for 3D Scene Graph Construction and Optimization. In: Robotics: Science and Systems XVIII. Robotics: Science and Systems Foundation (2022).https://doi. org/10.15607/RSS.2022.XVIII.050

    work page doi:10.15607/rss.2022.xviii.050 2022

  15. [15]

    In: Robotics: Science and Systems XIX

    Jatavallabhula, K., Kuwajerwala, A., Gu, Q., Omama, M., Iyer, G., Saryazdi, S., Chen, T., Maalouf, A., Li, S., Keetha, N., Tewari, A., Tenenbaum, J., Melo, C., Krishna, M., Paull, L., Shkurti, F., Torralba, A.: ConceptFusion: Open-set multi- Think While You Map 17 modal 3D mapping. In: Robotics: Science and Systems XIX. Robotics: Science and Systems Found...

    work page doi:10.15607/rss.2023.xix.066 2023

  16. [16]

    In: IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023

    Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: LERF: Lan- guage Embedded Radiance Fields. In: 2023 IEEE/CVF International Confer- ence on Computer Vision (ICCV). pp. 19672–19682. IEEE, Paris, France (2023). https://doi.org/10.1109/ICCV51070.2023.01807

    work page doi:10.1109/iccv51070.2023.01807 2023

  17. [17]

    In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Koch, S., Vaskevicius, N., Colosi, M., Hermosilla, P., Ropinski, T.: Open3dsg: Open-vocabulary 3d scene graphs from point clouds with queryable objects and open-set relationships. In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14183–14193 (2024).https://doi.org/10. 1109/CVPR52733.2024.01345

    work page arXiv 2024

  18. [18]

    IEEE Robotics and Automation Letters6(4), 7041–7048 (2021).https://doi.org/10.1109/LRA.2021.3097242

    Lin,S.,Wang,J.,Xu,M.,Zhao,H.,Chen,Z.:TopologyAwareObject-LevelSeman- tic Mapping Towards More Robust Loop Closure. IEEE Robotics and Automation Letters6(4), 7041–7048 (2021).https://doi.org/10.1109/LRA.2021.3097242

    work page doi:10.1109/lra.2021.3097242 2021

  19. [19]

    In: 2025 IEEE International Conference on Robotics and Automation (ICRA)

    Linok, S., Zemskova, T., Ladanova, S., Titkov, R., Yudin, D., Monastyrny, M., Valenkov, A.: Beyond Bare Queries: Open-Vocabulary Object Grounding with 3D Scene Graph. In: 2025 IEEE International Conference on Robotics and Automation (ICRA). pp. 13582–13589 (2025).https://doi.org/10.1109/ICRA55743.2025. 11128059

    work page doi:10.1109/icra55743.2025 2025

  20. [20]

    In: 2025 IEEE International Conference on Robotics and Automa- tion (ICRA)

    Liu, P., Guo, Z., Warke, M., Chintala, S., Paxton, C., Shafiullah, N.M.M., Pinto, L.: Dynamem: Online Dynamic Spatio-Semantic Memory for Open World Mobile Manipulation. In: 2025 IEEE International Conference on Robotics and Automa- tion (ICRA). pp. 13346–13355 (2025).https://doi.org/10.1109/ICRA55743. 2025.11127619

    work page doi:10.1109/icra55743 2025

  21. [21]

    In: Computer Vision – ECCV 2024, vol

    Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Jiang, Q., Li, C., Yang, J., Su, H., Zhu, J., Zhang, L.: Grounding DINO: Marrying DINO with Grounded Pre-training for Open-Set Object Detection. In: Computer Vision – ECCV 2024, vol. 15105, pp. 38–55. Springer Nature Switzerland, Cham (2025).https://doi. org/10.1007/978-3-031-72970-6_3

    work page doi:10.1007/978-3-031-72970-6_3 2024

  22. [22]

    In: 2019 International Conference on Robotics and Au- tomation (ICRA)

    Liu, Y., Petillot, Y., Lane, D., Wang, S.: Global Localization with Object-Level Semantics and Topology. In: 2019 International Conference on Robotics and Au- tomation (ICRA). pp. 4909–4915. IEEE, Montreal, QC, Canada (2019).https: //doi.org/10.1109/ICRA.2019.8794475

    work page doi:10.1109/icra.2019.8794475 2019

  23. [23]

    In: International Conference on Learning Representations (ICLR) (2023)

    Ma, X., Yong, S., Zheng, Z., Li, Q., Liang, Y., Zhu, S.C., Huang, S.: Sqa3d: Sit- uated question answering in 3d scenes. In: International Conference on Learning Representations (ICLR) (2023)

    2023

  24. [24]

    IEEE Robotics and Automation Letters9(10), 8921–8928 (2024)

    Maggio, D., Chang, Y., Hughes, N., Trang, M., Griffith, D., Dougherty, C., Cristo- falo, E., Schmid, L., Carlone, L.: Clio: Real-Time Task-Driven Open-Set 3D Scene Graphs. IEEE Robotics and Automation Letters9(10), 8921–8928 (2024). https://doi.org/10.1109/LRA.2024.3451395

    work page doi:10.1109/lra.2024.3451395 2024

  25. [25]

    In: 2018 International Conference on 3D Vision (3DV)

    McCormac, J., Clark, R., Bloesch, M., Davison, A., Leutenegger, S.: Fusion++: Volumetric object-level slam. In: 2018 International Conference on 3D Vision (3DV). pp. 32–41 (2018).https://doi.org/10.1109/3DV.2018.00015

    work page doi:10.1109/3dv.2018.00015 2018

  26. [26]

    In: International Conference on 3D Vision, 3DV 2025, Singapore, March 25- 28, 2025

    Mei, G., Riz, L., Wang, Y., Poiesi, F.: Vocabulary-Free 3D Instance Segmentation with Vision-Language Assistant. In: 2025 International Conference on 3D Vision (3DV). pp. 1197–1210 (2025).https://doi.org/10.1109/3DV66043.2025.00114

    work page doi:10.1109/3dv66043.2025.00114 2025

  27. [27]

    Meta: The Llama 4 herd: The beginning of a new era of natively multi- modal AI innovation (2025),https://ai.meta.com/blog/llama-4-multimodal- intelligence/

    2025

  28. [28]

    21276 18 D

    OpenAI: GPT-4o System Card (2024).https://doi.org/10.48550/arXiv.2410. 21276 18 D. Bickici et al

    work page doi:10.48550/arxiv.2410 2024

  29. [29]

    OpenAI: OpenAI GPT-5 System Card (2025).https://doi.org/10.48550/ arXiv.2601.03267

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  30. [30]

    OpenAI: Introducing GPT-5.4 (2026),https://openai.com/index/introducing- gpt-5-4/

    2026

  31. [31]

    Transactions on Ma- chine Learning Research (2024)

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Syn- naeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning robust visual fe...

    2024

  32. [32]

    In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: OpenScene: 3D Scene Understanding with Open Vocabularies. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 815–824. IEEE, Vancouver, BC, Canada (2023).https://doi.org/10.1109/CVPR52729. 2023.00085

    work page doi:10.1109/cvpr52729 2023

  33. [33]

    In: Robotics: Science and Systems (RSS) (2025).https://doi.org/10.15607/RSS.2025.XXI

    Peterson, M.B., Jia, Y.X., Tian, Y., Thomas, A., How, J.P.: Roman: Open-set object map alignment for robust view-invariant global localization. In: Robotics: Science and Systems (RSS) (2025).https://doi.org/10.15607/RSS.2025.XXI. 029

    work page doi:10.15607/rss.2025.xxi 2025

  34. [34]

    In: Proceedings of the 38th International Conference on Machine Learning (ICML)

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transfer- able visual models from natural language supervision. In: Proceedings of the 38th International Conference on Machine Learning (ICML). vol. 139, pp. 8748–8763. PMLR (2021)

    2021

  35. [35]

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., Mintun, E., Pan, J., Alwala, K.V., Carion, N., Wu, C.Y., Girshick, R., Dollár, P., Feichtenhofer, C.: SAM 2: Segment Anything in Images and Videos (2024).https://doi.org/10.48550/ARXIV.2408.00714

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2408.00714 2024

  36. [36]

    Reimers, N., Gurevych, I.: Sentence-BERT: Sentence embeddings using Siamese BERT-networks. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natu- ral Language Processing (EMNLP-IJCNLP). pp. 3982–3992. Association for Com- putational Linguistics, Hong Kong, China (2019)...

    2019

  37. [37]

    Grounded SAM: Assembling Open-World Models for Diverse Visual Tasks

    Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., Zeng, Z., Zhang, H., Li, F., Yang, J., Li, H., Jiang, Q., Zhang, L.: Grounded SAM: Assembling Open-World Models for Diverse Visual Tasks (2024). https://doi.org/10.48550/arXiv.2401.14159

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2401.14159 2024

  38. [38]

    In: 2025 International Conference on Machine Learning and Applications (ICMLA)

    Renz, M., Igelbrink, F., Atzmueller, M.: Integrating prior observations for incre- mental 3d scene graph prediction. In: 2025 International Conference on Machine Learning and Applications (ICMLA). pp. 887–892 (2025).https://doi.org/10. 1109/ICMLA66185.2025.00132

    work page arXiv 2025

  39. [39]

    In: IEEE Int

    Rosinol, A., Abate, M., Chang, Y., Carlone, L.: Kimera: an Open-Source Library for Real-Time Metric-Semantic Localization and Mapping. In: 2020 IEEE Inter- national Conference on Robotics and Automation (ICRA). pp. 1689–1696 (2020). https://doi.org/10.1109/ICRA40945.2020.9196885

    work page doi:10.1109/icra40945.2020.9196885 2020

  40. [40]

    The International Journal of Robotics Research40(12-14), 1510–1546 (2021).https://doi.org/10.1177/02783649211056674 Think While You Map 19

    Rosinol, A., Violette, A., Abate, M., Hughes, N., Chang, Y., Shi, J., Gupta, A., Carlone, L.: Kimera: From SLAM to spatial perception with 3D dynamic scene graphs. The International Journal of Robotics Research40(12-14), 1510–1546 (2021).https://doi.org/10.1177/02783649211056674 Think While You Map 19

    work page doi:10.1177/02783649211056674 2021

  41. [41]

    In: 2017 IEEE International Conference on Robotics and Au- tomation (ICRA)

    Runz, M., Agapito, L.: Co-fusion: Real-time segmentation, tracking and fusion of multiple objects. In: 2017 IEEE International Conference on Robotics and Au- tomation (ICRA). pp. 4471–4478. IEEE, Singapore, Singapore (2017).https: //doi.org/10.1109/ICRA.2017.7989518

    work page doi:10.1109/icra.2017.7989518 2017

  42. [42]

    In: 2018 IEEE International SymposiumonMixedandAugmentedReality(ISMAR).pp.10–20.IEEE,Munich, Germany (2018).https://doi.org/10.1109/ISMAR.2018.00024

    Runz, M., Buffier, M., Agapito, L.: MaskFusion: Real-Time Recognition, Track- ing and Reconstruction of Multiple Moving Objects. In: 2018 IEEE International SymposiumonMixedandAugmentedReality(ISMAR).pp.10–20.IEEE,Munich, Germany (2018).https://doi.org/10.1109/ISMAR.2018.00024

    work page doi:10.1109/ismar.2018.00024 2018

  43. [43]

    In: Robotics: Science and Systems XIX (2023).https://doi.org/10.15607/RSS.2023.XIX.074

    Shafiullah, N.M.M., Paxton, C., Pinto, L., Chintala, S., Szlam, A.: CLIP-Fields: Weakly Supervised Semantic Fields for Robotic Memory. In: Robotics: Science and Systems XIX (2023).https://doi.org/10.15607/RSS.2023.XIX.074

    work page doi:10.15607/rss.2023.xix.074 2023

  44. [44]

    Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur- Artal, R., Ren, C., Verma, S., Clarkson, A., Yan, M., Budge, B., Yan, Y., Pan, X., Yon, J., Zou, Y., Leon, K., Carter, N., Briales, J., Gillingham, T., Mueggler, E., Pesqueira, L., Savva, M., Batra, D., Strasdat, H.M., Nardi, R.D., Goesele, M., Lovegrove, S., Newcombe, R.:...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1906.05797 2019

  45. [45]

    In: Proceedings of the 37th International Conference on Neural Information Processing Systems (NeurIPS)

    Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: open-vocabulary 3d instance segmentation. In: Proceedings of the 37th International Conference on Neural Information Processing Systems (NeurIPS). Curran Associates Inc., Red Hook, NY, USA (2023)

    2023

  46. [46]

    In: 2025 IEEE/CVF International Con- ference on Computer Vision (ICCV)

    Wang, Z., Su, Y., Li, C., Wang, D., Huang, Y., Li, X., Zhao, B.: Open-vocabulary octree-graph for 3d scene understanding. In: 2025 IEEE/CVF International Con- ference on Computer Vision (ICCV). pp. 7037–7047 (2025).https://doi.org/10. 1109/ICCV51701.2025.00661

    work page arXiv 2025

  47. [47]

    In: Robotics: Science and Systems XX

    Werby, A., Huang, C., Büchner, M., Valada, A., Burgard, W.: Hierarchical Open- Vocabulary 3D Scene Graphs for Language-Grounded Robot Navigation. In: Robotics: Science and Systems XX. Robotics: Science and Systems Foundation (2024).https://doi.org/10.15607/RSS.2024.XX.077

    work page doi:10.15607/rss.2024.xx.077 2024

  48. [48]

    In���� �������� ���������� �� �������� ������ ��� ������� ����������� ������

    Wu, S.C., Tateno, K., Navab, N., Tombari, F.: Incremental 3D Semantic Scene Graph Prediction from RGB Sequences. In: 2023 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR). pp. 5064–5074. IEEE, Vancouver, BC, Canada (2023).https://doi.org/10.1109/CVPR52729.2023.00490

    work page doi:10.1109/cvpr52729.2023.00490 2023

  49. [49]

    In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021

    Wu, S.C., Wald, J., Tateno, K., Navab, N., Tombari, F.: SceneGraphFusion: Incre- mental 3D Scene Graph Prediction from RGB-D Sequences. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7511–7521. IEEE, Nashville, TN, USA (2021).https://doi.org/10.1109/CVPR46437.2021. 00743

    work page doi:10.1109/cvpr46437.2021 2021

  50. [51]

    In: 2024 IEEE International Conference on Robotics and Automa- tion (ICRA)

    Yang,J.,Chen,X.,Qian,S.,Madaan,N.,Iyengar,M.,Fouhey,D.F.,Chai,J.:LLM- Grounder: Open-Vocabulary 3D Visual Grounding with Large Language Model as an Agent. In: 2024 IEEE International Conference on Robotics and Automa- tion (ICRA). pp. 7694–7701 (2024).https://doi.org/10.1109/ICRA57147.2024. 10610443

    work page doi:10.1109/icra57147.2024 2024

  51. [52]

    Set-of-Mark Prompting Unleashes Extraordinary Visual Grounding in GPT-4V

    Yang, J., Zhang, H., Li, F., Zou, X., Li, C., Gao, J.: Set-of-Mark Prompting Un- leashes Extraordinary Visual Grounding in GPT-4V (2023).https://doi.org/ 10.48550/arXiv.2310.11441 20 D. Bickici et al

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2310.11441 2023

  52. [53]

    In: 2025 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, CVPR 2025, Nashville, TN, USA, June 11- 15, 2025

    Zhang, C., Delitzas, A., Wang, F., Zhang, R., Ji, X., Pollefeys, M., Engelmann, F.: Open-Vocabulary Functional 3D Scene Graphs for Real-World Indoor Spaces. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025).https://doi.org/10.1109/CVPR52734.2025.01807 Think While You Map 21 Think While You Map: Asynchronou...

    work page doi:10.1109/cvpr52734.2025.01807 2025

  53. [54]

    "label": A specific, fine-grained noun (e.g., ’winged armchair’ instead of ’chair’, ’ceramic vase’ instead of ’decor’). 30 D. Bickici et al

  54. [55]

    attributes

    "attributes": - "material": (e.g., wood, velvet, glass, plastic) - "color": (Dominant colors) - "state": (e.g., open, closed, folded, dirty, wet, empty) - "texture": (e.g., glossy, matte, rough, knitted) Constraint: - Do NOT hallucinate attributes not visible in the images. - If an object is completely unclear, set attributes to null. - Output pure JSON m...

  55. [56]

    Select the object ID that best matches the query

  56. [57]

    –- User –- Below is a 3D scene composed of objects and edges

    Provide a short, factual explanation. –- User –- Below is a 3D scene composed of objects and edges. Each edge is represented as [subject_id, predicate, object_id, distance_m (optional), compass (optional)]. {scene_line} {edge_line} query={utterance} I Full implementation details For completeness, Tab. 13 summarizes the key implementation parameters used i...