The DOI matches its Crossref record; 87 of the 87 references it cites were found in a registry or an index; 13 of the 17 sources its claims cite were checked for support.
ConceptGraphs: Open-Vocabulary 3D Scene Graphs for Perception and Planning
Open accessAbstract
For robots to perform a wide variety of tasks, they require a 3D representation of the world that is semantically rich, yet compact and efficient for task-driven perception and planning. Recent approaches have attempted to leverage features from large vision-language models to encode semantics in 3D representations. However, these approaches tend to produce maps with per-point feature vectors, which do not scale well in larger environments, nor do they contain semantic spatial relationships between entities in the environment, which are useful for downstream planning. In this work, we propose ConceptGraphs, an open-vocabulary graph-structured representation for 3D scenes. ConceptGraphs is built by leveraging 2D foundation models and fusing their output to 3D by multi-view association. The resulting representations generalize to novel semantic classes, without the need to collect large 3D datasets or finetune models. We demonstrate the utility of this representation through a number of downstream planning tasks that are specified through abstract (language) prompts and require complex reasoning over spatial and semantic concepts. To explore the full scope of our experiments and results, we encourage readers to visit our project webpage.
Overview
CoChat Summary
TL;DR
ConceptGraphs is an open-vocabulary, object-centric 3D scene graph built by fusing class-agnostic 2D segmentation and vision-language features into 3D objects, then using LVLMs/LLMs to caption nodes and infer edges. It supports language-queryable robot perception and planning across manipulation, navigation, localization, and map updates without 3D finetuning.
Key findings
- On Replica scene graph construction, ConceptGraphs achieved average node precision 0.71 and edge precision 0.88; the detector variant CG-D achieved 0.61 node precision and 0.91 edge precision.
- Open-vocabulary 3D semantic segmentation on Replica: ConceptGraphs reached 40.63 mAcc and 35.95 F-mIoU, compared with ConceptFusion+SAM at 31.53 mAcc and 38.70 F-mIoU.
- For text-query object retrieval, LLM-based retrieval handled complex queries better than CLIP: on Replica negation queries, CLIP R@1 was 0.26 vs LLM 0.80; on REAL Lab affordance queries, CLIP R@1 was 0.40 vs LLM 1.00; on REAL Lab negation queries, CLIP R@1 was 0.00 vs LLM 1.00.
- Real-robot demonstrations included a Clearpath Jackal UGV for navigation/object search/traversability and a Boston Dynamics Spot Arm for open-vocabulary pick-and-place, e.g., responding to “cuddly quacker” by grabbing a duck plush toy.
Methods
- Given posed RGB-D frames, ConceptGraphs runs class-agnostic segmentation (SAM in experiments), extracts visual descriptors (CLIP), projects masks to 3D, denoises with DBSCAN, and associates/fuses detections across views using geometric and semantic similarity.
- Each object node is captioned by an LVLM (LLaVA) from up to 10 best views, then refined by an LLM (GPT-4); scene graph edges are generated by computing 3D bounding-box IoU, pruning with a minimum spanning tree, and prompting an LLM to label spatial relationships.
- Evaluation used Replica and a real-world REAL Lab scan, with Amazon Mechanical Turk human evaluators for node/edge accuracy, plus semantic segmentation, text-query retrieval, and real-robot tasks on wheeled and legged platforms.
Limitations
- The paper states node captioning errors from LVLMs like LLaVA, occasional missed small/thin objects, duplicate detections, and significant computational/economic costs from multiple LVLM and proprietary LLM inferences.
- The appendix adds that LLaVA-7B misclassified a non-negligible number of small objects as toothbrushes or pairs of scissors, and that the work does not explicitly focus on improving LLM-based planning over 3D scene graphs.
- Evident from the text: open-vocabulary node/edge quality was evaluated via human evaluators rather than automated metrics, and LLM retrieval performance is limited by node caption accuracy.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
4 Valid · 7 Not valid · 2 Not checked · 30 Own results and interpretations
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“First, node captioning incurs errors due to the current limitations of LVLMs like LLaVA [55].”
What the sources say
- Could not verifycites [55] Zhang, 2023, Recognize anything: A strong image tagging model sourceExperimental — accuracy not yet measured
The “Compared with Vision-Language Models” section discusses accuracy limitations of CLIP and BLIP, and the Figure 2 caption says BLIP’s tag recall is limited by caption generation. Neither passage addresses node-captioning errors or attributes them to limitations of LLaVA; searches for LLaVA and captioning errors found no passage that does. (LLM, conf=0.75)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on an assumption not shown here.
- Could not verify
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“By using an LVLM [55] to add a description of the current camera image to the text prompt, the robot can also answer visual queries.”
What the sources say
- Partly supportedcites [55] Zhang, 2023, Recognize anything: A strong image tagging model sourceExperimental — accuracy not yet measured
The Model Architecture section (page unavailable) says RAM includes a text-generation encoder-decoder for image captioning, supporting the idea that it can describe an image. It does not describe adding a camera-image description to a robot’s text prompt or answering visual queries, so those parts of the attributed claim are unsupported. (LLM, conf=0.82)
plainstates a causeSingle case- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“However, these approaches tend to produce maps with per-point feature vectors, which do not scale well in larger environments”
- Also stated in: § A. Related Work · p. 2 · plain
“First, assigning every point a semantic feature vector is highly redundant and consumes more memory than necessary, greatly limiting scalability to large scenes.”
What the sources say
- Partly supportedcites [17] sourceExperimental — accuracy not yet measured
The Limitations paragraph (VII. Conclusion, p. 10) says ConceptFusion stores high-dimensional embeddings for millions of points in apartment-scale maps, requiring large amounts of memory. This supports a scaling concern for ConceptFusion, but the paper does not establish that recent approaches generally use per-point vectors or scale poorly in larger environments. (LLM, conf=0.84)
- Partly supportedcites [18] sourceExperimental — accuracy not yet measured
The Method section describes a per-point feature vector for every 3D point. The implementation details also report sampling only 20,000 point features per training iteration because of memory constraints. Our copy does not establish the broader claim that maps using these features do not scale well in larger environments. (LLM, conf=0.85)
- Partly supportedcites [19] sourceExperimental — accuracy not yet measured
The Related Work section (page not available in our copy) says recent 3D approaches align points with image regions to distill dense semantic embeddings and may require heavy disk storage. It does not say these approaches produce maps with per-point feature vectors or establish that they fail to scale in larger environments. (LLM, conf=0.83)
- Partly supportedcites [20] sourceExperimental — accuracy not yet measured
Section 3.2.1 describes projected features for individual points, supporting the per-point feature-vector part of the claim. The Introduction discusses computation and memory costs for multi-view projection, but our copy does not say that storing per-point feature maps fails to scale in larger environments. The paper gives a different reason for the costs it identifies. (LLM, conf=0.79)
- Partly supportedcites [21] sourceExperimental — accuracy not yet measured
Section 3.2 (p. 4) describes a feature vector for each point in a 3D scene. Section 4.2 (p. 5) says inferring point semantics by projecting features from every RGB view is very slow and impractical. The paper does not say that storing per-point feature vectors fails to scale in larger environments; its stated speed concern is about feature projection. (LLM, conf=0.85)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § A. Related Work · p. 2 · plain
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“In particular, the LLaVA-7B [55] model used for node captioning misclassifies a non-negligible number of small objects as toothbrushes or pairs of scissors .”
What the sources say
- Could not verifycites [55] Zhang, 2023, Recognize anything: A strong image tagging model sourceExperimental — accuracy not yet measured
The Model Ablation Study (section 4.3; page unavailable) discusses RAM’s tagging ability, while the comparison with vision-language models discusses CLIP and BLIP. Searches for LLaVA-7B, toothbrushes, scissors, and small-object misclassification found no passage addressing the claimed captioning errors. (LLM, conf=0.76)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“For example, when shown a picture of Michael Jordan and prompted with Something this guy would play with , the robot finds a basketball.”
plainstates a causeSingle case- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
9 more claims
Create a free account to see all 43 claims and to verify the 2 references still open.
- It may state two separate claims as one. 21
- Probably not a substantial claim. 10
- The quote only partly states this claim. 1
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Nonetheless, thanks to the compact graph structure, ConceptGraphs and HOV-SG are much smaller than their dense counterparts.”
From Hierarchical Open-Vocabulary 3D Scene Graphs for Language-Grounded Robot Navigation (2024) · cites this paper as [14]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“HOV-SG extends the scope of potential navigation goals to more specific spatial concepts such as regions and floors compared to simple object goals [14, 16, 10, 13].”
From Hierarchical Open-Vocabulary 3D Scene Graphs for Language-Grounded Robot Navigation (2024) · cites this paper as [14]
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“As shown in Table V, HOV-SG shows a significant performance increase of 11.69% on object-room-floor queries and a 2.2% advantage on object-room queries when compared with ConceptGraphs.”
From Hierarchical Open-Vocabulary 3D Scene Graphs for Language-Grounded Robot Navigation (2024) · cites this paper as [14]
The cited source says what the claim attributes to it.
In the citing paper
“As shown in Table V, HOV-SG shows a significant performance increase of 11.69% on object-room-floor queries and a 2.2% advantage on object-room queries when compared with ConceptGraphs.”
From Hierarchical Open-Vocabulary 3D Scene Graphs for Language-Grounded Robot Navigation (2024) · cites this paper as [14]
The cited source says what the claim attributes to it.
In the citing paper
“Their semantic understanding and reasoning capabilities have been increasingly applied in robotics-for example, in contextual navigation [8], [9], affordance-driven grasping and manipulation [10], [11], and guiding reward design or data collection for skill learning [12].”
From SwarmGPT: Combining Large Language Models With Safe Motion Planning for Drone Swarm Choreography (2025) · cites this paper as [8]
2 more claims about this paper
Create a free account to see every claim other papers make about this one.
Citations
Cites
High-Resolution Image Synthesis with Latent Diffusion Models
Rombach, R., Blattmann, A., Lorenz, D., Esser, P., & Ommer, B. (2022) · 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr52688.2022.01042 · Cited by 20,302 · cited 1×
“There have been significant recent efforts [16]–[30] focused on building 3D representations by leveraging foundation models - large, powerful models that capture a diverse set of concepts and accomplish a wide range of tasks [31]– [35].”
BackgroundLearning Transferable Visual Models From Natural Language Supervision
Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. (2021) · arXiv · doi:10.48550/arxiv.2103.00020 · Cited by 15,293
Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A. C., Lo, W.-Y., et al. (2023) · 2023 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv51070.2023.00371 · Cited by 11,953 · cited 3×
“At the core of ConceptGraphs is an object-centric 3D mapping technique that integrates geometric cues from conventional 3D mapping systems, and semantic cues from vision and language foundation models [31], [33], [34], [54]–[56].”
MethodStructure-from-Motion Revisited
Schonberger, J. L., & Frahm, J.-M. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.445 · Cited by 6,433 · cited 1×
“Closed-vocabulary semantic mapping in 3D. Early works reconstruct the 3D map through online algorithms like simultaneous localization and mapping (SLAM) [1]–[5] or offline methods like structure-from-motion (SfM) [6], [7].”
BackgroundKinectFusion: Real-time dense surface mapping and tracking
Newcombe, R. A., Fitzgibbon, A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A. J., Kohi, P., Shotton, J., & Hodges, S. (2011) · 2011 10th IEEE International Symposium on Mixed and Augmented Reality · doi:10.1109/ismar.2011.6092378 · Cited by 4,008 · cited 1×
“Closed-vocabulary semantic mapping in 3D. Early works reconstruct the 3D map through online algorithms like simultaneous localization and mapping (SLAM) [1]–[5] or offline methods like structure-from-motion (SfM) [6], [7].”
BackgroundPixelwise View Selection for Unstructured Multi-View Stereo
Schönberger, J. L., Zheng, E., Frahm, J.-M., & Pollefeys, M. (2016) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-46487-9_31 · Cited by 2,315 · cited 1×
“Early works reconstruct the 3D map through online algorithms like simultaneous localization and mapping (SLAM) [1]–[5] or offline methods like structure-from-motion (SfM) [6], [7].”
BackgroundDTAM: Dense tracking and mapping in real-time
Newcombe, R. A., Lovegrove, S. J., & Davison, A. J. (2011) · 2011 International Conference on Computer Vision · doi:10.1109/iccv.2011.6126513 · Cited by 2,078
Grounding DINO: Marrying DINO with Grounded Pre-training for Open-Set Object Detection
Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Jiang, Q., Li, C., Yang, J., Su, H., et al. (2025) · Lecture Notes in Computer Science · doi:10.1007/978-3-031-72970-6_3 · Cited by 2,074
DINOv2: Learning Robust Visual Features without Supervision
Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al. (2023) · arXiv · doi:10.48550/arxiv.2304.07193 · Cited by 1,078 · cited 2×
“Each extracted mask m t,i is then passed to a visual feature extractor (CLIP [31], DINO [54]) to obtain a visual descriptor f t,i = Embed ( I rgb t , m t,i ) .”
MethodElasticFusion: Dense SLAM Without A Pose Graph
Whelan, T., Leutenegger, S., Salas Moreno, R., Glocker, B., & Davison, A. (2015) · Robotics: Science and Systems XI · doi:10.15607/rss.2015.xi.001 · Cited by 835
Research works printed but not linked: 1
- Shen, Distilled feature fields enable few-shot manipulation
Cited by
A survey on integration of large language models with intelligent robots
Kim, Y., Kim, D., Choi, J., Park, J., Oh, N., & Park, D. (2024) · Intelligent Service Robotics · doi:10.1007/s11370-024-00550-5 · Cited by 234
Aligning Cyber Space With Physical World: A Comprehensive Survey on Embodied AI
Liu, Y., Chen, W., Bai, Y., Liang, X., Li, G., Gao, W., & Lin, L. (2025) · IEEE/ASME Transactions on Mechatronics · doi:10.1109/tmech.2025.3574943 · Cited by 521
“Similarly, ConceptGraphs [189] also adopts 3-”
Hierarchical Open-Vocabulary 3D Scene Graphs for Language-Grounded Robot Navigation
Werby, A., Huang, C., Büchner, M., Valada, A., & Burgard, W. (2024) · Robotics: Science and Systems XX · doi:10.15607/rss.2024.xx.077 · Cited by 124
“This is crucial as room categories are not always fully determinable when labeling, e.g., combined kitchen and living room areas.”
BackgroundMethodInfluentialLanguage-Grounded Dynamic Scene Graphs for Interactive Object Search With Mobile Manipulation
Honerkamp, D., Büchner, M., Despinoy, F., Welschehold, T., & Valada, A. (2024) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2024.3441495 · Cited by 157
“As a result, simple prompting strategies, such as lists of observed objects [13], [16], [19] or raw JSON input [5] of a full scene graph to a language model becomes insufficient, as we demonstrate in our experiments.”
BackgroundMethodOpen3DIS: Open-Vocabulary 3D Instance Segmentation with 2D Mask Guidance
Nguyen, P., Ngo, T., Kalogerakis, E., Gan, C., Tran, A. T., Pham, C. H., & Nguyen, K. (2024) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr52733.2024.00385 · Cited by 97
“This capability is highlighted in recent work [12, 17, 37] for OV-3DSS and [3, 34, 69] for OV-3DOD.”
BackgroundSelf-citation
Metadata
- Authors
- 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.
- Affiliations
- University of Toronto (CA)
- Université de Montréal
- Université de Montréal (CA)
- MIT
- Johns Hopkins University Applied Physics Laboratory (US)
- Johns Hopkins University (US)
- UMass
- DEVCOM Army Research Laboratory (US)
- Massachusetts Institute of Technology (US)
- ORCID
- Venue
- 2024 IEEE International Conference on Robotics and Automation (ICRA)
- Publisher
- IEEE
- Published
- 2024-05-13
- Pages
- 5021-5028
- ISSN
- 1050-4729 (ISSN-L)
- OpenAlex
- W4401414715
- Semantic Scholar
- 174a3290ff0040e0f6a7a0b43bcd752c456350a0
- Cited by
- 406
- References
- 80
- Citations by year
- 2026: 210 · 2025: 131 · 2024: 62
- FWCI
- 34.97
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- Proceedings - IEEE International Conference on Robotics and Automation/Proceedings (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Funders
- Army Research Laboratory
- Topics
- Physical Sciences › Computer Science › Computer Graphics and Computer-Aided Design › 3D Shape Modeling and Analysis
- Physical Sciences › Engineering › Control and Systems Engineering › Robot Manipulation and Learning
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Multimodal Machine Learning Applications
- Keywords
- 3D scene graphs, robot perception, robot planning, vision-language models, 3D scene understanding, spatial reasoning, language prompts
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-11-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-11-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: ConceptGraphs: Open-Vocabulary 3D Scene Graphs for Perception and Planning. https://app.cochat.ai/papers/10.1109/icra57147.2024.10610243
- The paper
- 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. (2024). ConceptGraphs: Open-Vocabulary 3D Scene Graphs for Perception and Planning. 2024 IEEE International Conference on Robotics and Automation (ICRA). https://doi.org/10.1109/icra57147.2024.10610243
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.