pith. machine review for the scientific record. sign in

arxiv: 2604.04108 · v1 · submitted 2026-04-05 · 💻 cs.CV

Recognition: no theorem link

Hypothesis Graph Refinement: Hypothesis-Driven Exploration with Cascade Error Correction for Embodied Navigation

Peixin Chen , Guoxi Zhang , Jianwei Ma , Qing Li
Authors on Pith no claims yet

Pith reviewed 2026-05-13 17:02 UTC · model grok-4.3

classification 💻 cs.CV
keywords hypothesis graph refinementembodied navigationcascade error correctionfrontier explorationgraph memorylifelong navigationvision-language modelssemantic prediction
0
0 comments X

The pith

Representing frontier predictions as revisable hypothesis nodes in a dependency graph allows embodied agents to retract semantic errors by pruning entire dependent subgraphs upon mismatch.

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

Embodied agents exploring unknown spaces build maps from partial observations but risk embedding wrong predictions about unseen areas that then mislead later decisions. The paper shows how treating those predictions as nodes in a graph with explicit dependencies enables a correction process that removes not only the contradicted node but all nodes that relied on it. This contraction keeps the stored structure accurate over long episodes instead of letting additive errors compound. The result is more directed exploration and higher success on navigation and question-answering tasks in simulated lifelong settings.

Core claim

Hypothesis Graph Refinement represents frontier predictions as revisable hypothesis nodes inside a dependency-aware graph memory and applies verification-driven cascade correction that retracts any refuted node together with all its downstream dependents once on-site observations contradict the predicted semantics.

What carries the argument

Verification-driven cascade correction, which compares new observations against stored semantic predictions and prunes the refuted hypothesis node along with every dependent node that was built upon it.

If this is right

  • Yields 72.41 percent success rate and 56.22 percent SPL on GOAT-Bench multimodal lifelong navigation.
  • Eliminates roughly 20 percent of structurally redundant hypothesis nodes through pruning.
  • Cuts revisits to erroneous regions by a factor of 4.5 compared with baselines.
  • Produces consistent gains on the A-EQA and EM-EQA embodied question-answering benchmarks.

Where Pith is reading between the lines

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

  • Dependency tracking of this kind could transfer to other sequential decision settings where partial observations generate assumptions that later need targeted rollback.
  • The method implies that memory structures in long-horizon agents benefit more from explicit retraction rules than from simple confidence decay.
  • Real-world deployment would need to test whether sensor noise still allows the mismatch detection step to trigger corrections at the right times.

Load-bearing premise

That on-site observations can reliably detect when a semantic prediction is wrong and that removing the dependent nodes discards only erroneous structure without losing still-valid information.

What would settle it

An experiment in which cascade correction produces lower success rates than a version without pruning, for instance because valid paths are removed more often than mistaken ones, would show the mechanism does not improve reliability.

Figures

Figures reproduced from arXiv: 2604.04108 by Guoxi Zhang, Jianwei Ma, Peixin Chen, Qing Li.

Figure 1
Figure 1. Figure 1: Overview of Hypothesis Graph Refinement (HGR). [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of HGR. (Left) Frontiers Ft detected from the occupancy map are fed to a VLM reasoner for semantic hypothesis module, which estimates categorical distributions and generates hypothesis nodes linked to observed nodes via spatial and dependency edges. Upon visitation, cascade correction compares predicted (Ipred) and actual (Iactual) semantics; if ∆sem > θ, the refuted node and all its dependent… view at source ↗
Figure 3
Figure 3. Figure 3: Semantic Hypothesis Module. Left: Traditional frontier representation treats unexplored regions as undifferentiated boundaries. Right: HGR projects proba￾bilistic semantic distributions onto frontiers as hypothesis nodes, enabling goal-directed exploration. vation, and the node transitions to V obs with updated labels; (2) Refutation— significant deviation triggers cascade correction, removing the node and… view at source ↗
Figure 4
Figure 4. Figure 4: Cascade Correction Example. A VLM misidentifies a mirror reflection as a bedroom entrance, generating hypothesis nodes for inferred furniture. Upon reaching the mirror and detecting a prediction violation (residual > θrefute), the system traces the dependency DAG and removes the entire erroneous subgraph, including all descendant hypothesis nodes. gation goal g: S(f_j; g) = \underbrace {P(c_g \mid \mathcal… view at source ↗
Figure 5
Figure 5. Figure 5: Cumulative Success Rate vs. Episode Steps. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative Comparison: Mirror-Induced Prediction Error. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative Failure Cases of HGR. We visualize four typical failure pat￾terns: (a) Object confusion caused by detector errors and lack of feature extraction; (b) Loss of visual details due to long observation distances; (c) State misjudgment caused by suboptimal viewpoints or missing subtle visual cues; and (d) Misunderstanding of spatial predicates (e.g., “between”) [PITH_FULL_IMAGE:figures/full_fig_p028… view at source ↗
read the original abstract

Embodied agents must explore partially observed environments while maintaining reliable long-horizon memory. Existing graph-based navigation systems improve scalability, but they often treat unexplored regions as semantically unknown, leading to inefficient frontier search. Although vision-language models (VLMs) can predict frontier semantics, erroneous predictions may be embedded into memory and propagate through downstream inferences, causing structural error accumulation that confidence attenuation alone cannot resolve. These observations call for a framework that can leverage semantic predictions for directed exploration while systematically retracting errors once new evidence contradicts them. We propose Hypothesis Graph Refinement (HGR), a framework that represents frontier predictions as revisable hypothesis nodes in a dependency-aware graph memory. HGR introduces (1) semantic hypothesis module, which estimates context-conditioned semantic distributions over frontiers and ranks exploration targets by goal relevance, travel cost, and uncertainty, and (2) verification-driven cascade correction, which compares on-site observations against predicted semantics and, upon mismatch, retracts the refuted node together with all its downstream dependents. Unlike additive map-building, this allows the graph to contract by pruning erroneous subgraphs, keeping memory reliable throughout long episodes. We evaluate HGR on multimodal lifelong navigation (GOAT-Bench) and embodied question answering (A-EQA, EM-EQA). HGR achieves 72.41% success rate and 56.22% SPL on GOAT-Bench, and shows consistent improvements on both QA benchmarks. Diagnostic analysis reveals that cascade correction eliminates approximately 20% of structurally redundant hypothesis nodes and reduces revisits to erroneous regions by 4.5x, with specular and transparent surfaces accounting for 67% of corrected prediction errors.

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 / 1 minor

Summary. The paper proposes Hypothesis Graph Refinement (HGR), a framework for embodied navigation in partially observed environments. It represents frontier predictions as revisable hypothesis nodes within a dependency-aware graph memory. The method introduces a semantic hypothesis module that estimates context-conditioned semantic distributions over frontiers and ranks targets by goal relevance, travel cost, and uncertainty, along with a verification-driven cascade correction mechanism that retracts a refuted hypothesis node and all its downstream dependents upon mismatch between on-site RGB-D observations and predicted semantics. Unlike additive map-building, the graph contracts by pruning erroneous subgraphs. Evaluation on GOAT-Bench reports 72.41% success rate and 56.22% SPL, with consistent gains on A-EQA and EM-EQA; diagnostics indicate cascade correction removes ~20% redundant nodes, reduces erroneous revisits by 4.5x, and corrects 67% of errors from specular/transparent surfaces.

Significance. If the cascade correction reliably identifies mismatches and prunes only erroneous structure without discarding valid information, HGR could meaningfully advance long-horizon embodied navigation by enabling directed exploration while maintaining memory reliability. The reported 4.5x reduction in erroneous revisits and performance on GOAT-Bench, A-EQA, and EM-EQA suggest practical efficiency gains over standard graph-based systems that rely on confidence attenuation alone. The diagnostic breakdown of error sources (specular/transparent surfaces) provides useful insight into VLM limitations in navigation.

major comments (3)
  1. The central performance claims (72.41% SR and 56.22% SPL on GOAT-Bench) are presented without baseline comparisons, error bars, statistical tests, or details on evaluation protocols, data exclusion criteria, or how success/SPL are computed, preventing assessment of whether the numbers support the superiority of cascade correction over prior methods.
  2. The verification-driven cascade correction is load-bearing for the reliability claim, yet the manuscript provides no quantitative evaluation (precision, recall, or failure modes) of the mismatch detection step between VLM predictions and on-site observations under realistic conditions such as sensor noise, partial views, or lighting variation, despite noting that specular/transparent surfaces cause 67% of corrected errors.
  3. The assumption that retracting a refuted node plus all downstream dependents removes only erroneous structure (without discarding still-valid frontier hypotheses) is not validated; the reported 4.5x reduction in erroneous revisits and ~20% node elimination rest directly on this untested property of the dependency graph.
minor comments (1)
  1. The abstract and diagnostic analysis would benefit from explicit definition of the dependency graph construction and how acyclicity is enforced to ensure cascade retraction is well-defined.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed review. The comments highlight important areas for strengthening the evaluation and validation of Hypothesis Graph Refinement. We address each major comment below and will incorporate the requested additions and clarifications in the revised manuscript.

read point-by-point responses

  1. Referee: The central performance claims (72.41% SR and 56.22% SPL on GOAT-Bench) are presented without baseline comparisons, error bars, statistical tests, or details on evaluation protocols, data exclusion criteria, or how success/SPL are computed, preventing assessment of whether the numbers support the superiority of cascade correction over prior methods.

    Authors: We agree that the current presentation of results lacks sufficient context for rigorous comparison. In the revision we will add a dedicated baselines subsection comparing HGR against frontier-based exploration, standard graph memory without hypothesis nodes, and recent VLM-driven navigation methods. Results will be reported with standard error bars across multiple random seeds, accompanied by statistical significance tests (paired t-tests and Wilcoxon signed-rank). A new appendix will detail the exact computation of success rate and SPL, the GOAT-Bench evaluation protocol, episode termination criteria, and any data exclusion rules applied. revision: yes

  2. Referee: The verification-driven cascade correction is load-bearing for the reliability claim, yet the manuscript provides no quantitative evaluation (precision, recall, or failure modes) of the mismatch detection step between VLM predictions and on-site observations under realistic conditions such as sensor noise, partial views, or lighting variation, despite noting that specular/transparent surfaces cause 67% of corrected errors.

    Authors: We acknowledge the absence of quantitative metrics for the mismatch detection module. The revised manuscript will include a new diagnostic subsection that reports precision, recall, and F1-score for the verification step. We will add controlled experiments that inject sensor noise, simulate partial views, and vary lighting conditions, together with a tabulated breakdown of failure modes. The 67% attribution to specular and transparent surfaces will be supported by per-scene counts and representative RGB-D examples showing both successful and unsuccessful detections. revision: yes

  3. Referee: The assumption that retracting a refuted node plus all downstream dependents removes only erroneous structure (without discarding still-valid frontier hypotheses) is not validated; the reported 4.5x reduction in erroneous revisits and ~20% node elimination rest directly on this untested property of the dependency graph.

    Authors: We agree that direct validation of pruning selectivity is required. In the revision we will add an analysis that tracks each pruned node and determines whether it would have produced an erroneous revisit if retained (via oracle re-evaluation on held-out trajectories). We will report the fraction of pruned nodes that were verifiably incorrect versus those that were still potentially valid, and include qualitative visualizations of the dependency graph before and after cascade correction to illustrate preservation of independent valid frontiers. These additions will ground the 4.5x and 20% figures in explicit selectivity measurements. revision: yes

Circularity Check

0 steps flagged

No circularity: method and results are self-contained with external benchmark evaluation

full rationale

The paper defines HGR procedurally as a graph-based framework with a semantic hypothesis module for ranking frontiers and a verification-driven cascade correction for retracting mismatched nodes plus dependents. Reported metrics (72.41% SR, 56.22% SPL on GOAT-Bench) are obtained via direct evaluation on independent external benchmarks (GOAT-Bench, A-EQA, EM-EQA) rather than any derivation that reduces performance to fitted parameters or self-referential definitions. No equations, self-citations, or ansatzes are invoked that would make the central claims equivalent to their inputs by construction. The derivation chain remains independent of the reported outcomes.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 2 invented entities

The framework rests on domain assumptions about the utility of VLM semantic predictions and the reliability of on-site mismatch detection; it introduces two new conceptual entities (hypothesis nodes and cascade correction) without independent falsifiable evidence outside the empirical results.

axioms (2)
  • domain assumption Vision-language models produce context-conditioned semantic distributions over frontiers that are sufficiently accurate to rank exploration targets usefully
    Invoked by the semantic hypothesis module for ranking by goal relevance, cost, and uncertainty.
  • domain assumption On-site observations can be compared directly against predicted semantics to detect contradictions reliably
    Required for the verification-driven cascade correction step.
invented entities (2)
  • Hypothesis nodes no independent evidence
    purpose: Represent revisable frontier semantic predictions inside a dependency-aware graph memory
    Core new representation that enables later retraction; no independent evidence provided beyond the method description.
  • Cascade correction mechanism no independent evidence
    purpose: Retract a refuted hypothesis node together with all its downstream dependents to contract the graph
    Central innovation for preventing structural error accumulation; no independent evidence outside the reported diagnostics.

pith-pipeline@v0.9.0 · 5604 in / 1567 out tokens · 58924 ms · 2026-05-13T17:02:32.720375+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

39 extracted references · 39 canonical work pages · 1 internal anchor

  1. [1]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    An, D., Wang, H., Wang, W., Wang, Z., Huang, Y., He, K., Wang, L.: Etpnav: Evolving topological planning for vision-language navigation in continuous envi- ronments. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    work page 2024

  2. [2]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I., Gould, S., Van Den Hengel, A.: Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3674–3683 (2018)

    work page 2018

  3. [3]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Armeni, I., He, Z.Y., Gwak, J., Zamir, A.R., Fischer, M., Malik, J., Savarese, S.: 3d scene graph: A structure for unified semantics, 3d space, and camera. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5664–5673 (2019)

    work page 2019

  4. [4]

    In: Findings of the Association for Computational Linguistics: EMNLP 2025 (2025)

    Chakraborty, T., Ghosh, U., Zhang, X., Niloy, F.F., Dong, Y., Li, J., Roy- Chowdhury, A.K., Song, C.: HEAL: An empirical study on hallucinations in em- bodied agents driven by large language models. In: Findings of the Association for Computational Linguistics: EMNLP 2025 (2025)

    work page 2025

  5. [5]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Deng, Z., Narasimhan, K., Russakovsky, O.: Evolving graphical planner: Contex- tual global planning for vision-and-language navigation. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 33 (2020)

    work page 2020

  6. [6]

    Interleave-VLA: Enhancing robot manipulation with interleaved image-text instructions.arXiv preprint arXiv:2505.02152, 2025

    Fan, C., Jia, X., Sun, Y., Wang, Y., Wei, J., Gong, Z., Zhao, X., Tomizuka, M., Yang, X., Yan, J., Ding, M.: Interleave-VLA: Enhancing robot manipulation with interleaved image-text instructions. arXiv preprint arXiv:2505.02152 (2025)

    work page arXiv 2025

  7. [7]

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

    Gadre, S.Y., Wortsman, M., Ilharco, G., Schmidt, L., Song, S.: CoWs on Pas- ture: Baselines and benchmarks for language-driven zero-shot object navigation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)

    work page 2023

  8. [8]

    In: Proceedings of the IEEE Inter- national Conference on Robotics and Automation (ICRA) (2024)

    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., Tenen- baum, J.B., Torralba, A., Shkurti, F., Paull, L.: ConceptGraphs: Open-vocabulary 3d scene graphs for perception and planning. In: Proceedings of the IEEE Inter- national Conference on Robotics and A...

    work page 2024

  9. [9]

    Heo, K., Kim, G., Kim, S., Cho, M.: Object-centric representation learning for enhanced3dscenegraphprediction.In:AdvancesinNeuralInformationProcessing Systems (NeurIPS) (2025)

    work page 2025

  10. [10]

    In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition

    Hong, Y., Wu, Q., Qi, Y., Rodriguez-Opazo, C., Gould, S.: Vln bert: A recur- rent vision-and-language bert for navigation. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 1643–1653 (2021)

    work page 2021

  11. [11]

    arXiv preprint arXiv:2505.07868 (2025)

    Huang, Y., Wu, M., Li, R., Tu, Z.: VISTA: Generative visual imagination for vision-and-language navigation. arXiv preprint arXiv:2505.07868 (2025)

    work page arXiv 2025

  12. [12]

    Language Models (Mostly) Know What They Know

    Kadavath, S., Conerly, T., Askell, A., Henighan, T., Drain, D., Perez, E., Schiefer, N., Hatfield-Dodds, Z., DasSarma, N., Tran-Johnson, E., et al.: Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221 (2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  13. [13]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Khanna, M., Ramrakhya, R., Chhablani, G., Yenamandra, S., Gervet, T., Chang, M., Kira, Z., Chaplot, D.S., Batra, D., Mottaghi, R.: GOAT-Bench: A benchmark for multi-modal lifelong navigation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16373–16383 (2024)

    work page 2024

  14. [14]

    In: 16 P

    Ku, A., Anderson, P., Patel, R., Ie, E., Baldridge, J.: Room-across-room: Multi- lingual vision-and-language navigation with dense spatiotemporal grounding. In: 16 P. Chen et al. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 4392–4412 (2020)

    work page 2020

  15. [15]

    In: Proceedings of the 40th International Conference on Machine Learning

    Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: bootstrapping language-image pre- training with frozen image encoders and large language models. In: Proceedings of the 40th International Conference on Machine Learning. ICML’23, JMLR.org (2023)

    work page 2023

  16. [16]

    arXiv preprint arXiv:2301.02382 (2023)

    Liu, J., Guo, J., Meng, Z., Xue, J.: Revolt: Relational reasoning and voronoi lo- cal graph planning for target-driven navigation. arXiv preprint arXiv:2301.02382 (2023)

    work page arXiv 2023

  17. [17]

    arXiv preprint arXiv:2409.15658 (2024)

    Liu, S., Du, J., Xiang, S., Wang, Z., Luo, D.: Long-horizon embodied plan- ning with implicit logical inference and hallucination mitigation. arXiv preprint arXiv:2409.15658 (2024)

    work page arXiv 2024

  18. [18]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2022)

    Majumdar, A., Aggarwal, G., Devnani, B., Hoffman, J., Batra, D.: ZSON: Zero- shot object-goal navigation using multimodal goal embeddings. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)

    work page 2022

  19. [19]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Majumdar, A., Ajay, A., Zhang, X., Putta, P., Yenamandra, S., Henaff, M., Silwal, S., Mcvay, P., Maksymets, O., Arnaud, S., Yadav, K., Li, Q., Newman, B., Sharma, M., Berges, V., Zhang, S., Agrawal, P., Bisk, Y., Batra, D., Kalakrishnan, M., Meier, F., Paxton, C., Sax, A., Rajeswaran, A.: OpenEQA: Embodied question answering in the era of foundation model...

    work page 2024

  20. [20]

    In: Proceedings of the 2023ConferenceonEmpiricalMethodsinNaturalLanguageProcessing(EMNLP)

    Manakul, P., Liusie, A., Gales, M.J.F.: SelfCheckGPT: Zero-resource black-box hallucination detection for generative large language models. In: Proceedings of the 2023ConferenceonEmpiricalMethodsinNaturalLanguageProcessing(EMNLP). pp. 9004–9017 (2023)

    work page 2023

  21. [21]

    Film: Following instructions in language with modular methods.arXiv preprint arXiv:2110.07342, 2021

    Min, S.Y., Chaplot, D.S., Ravikumar, P., Bisk, Y., Salakhutdinov, R.: Film: Following instructions in language with modular methods. arXiv preprint arXiv:2110.07342 (2021)

    work page arXiv 2021

  22. [22]

    Monaci, G., de Rezende, R.S., Deffayet, R., Csurka, G., Bono, G., D’ejean, H., Clin- chant, S., Wolf, C.: Rana: Retrieval-augmented navigation. Trans. Mach. Learn. Res.2025(2025)

    work page 2025

  23. [23]

    In: Proceedings of the 41st International Conference on Machine Learning

    Nasiriany, S., Xia, F., Yu, W., Xiao, T., Liang, J., Dasgupta, I., Xie, A., Driess, D., Wahid, A., Xu, Z., Vuong, Q., Zhang, T., Lee, T.W.E., Lee, K.H., Xu, P., Kirmani, S., Zhu, Y., Zeng, A., Hausman, K., Heess, N., Finn, C., Levine, S., Ichter, B.: PIVOT: Iterative visual prompting elicits actionable knowledge for VLMs. In: Proceedings of the 41st Inter...

    work page 2024

  24. [24]

    Vln-r1: Vision-language navigation via reinforcement fine-tuning.arXiv preprint arXiv:2506.17221,

    Qi, Z., Zhang, Z., Yu, Y., Wang, J., Zhao, H.: VLN-R1: Vision-language navigation via reinforcement fine-tuning. arXiv preprint arXiv:2506.17221 (2025)

    work page arXiv 2025

  25. [25]

    In: Robotics: Science and Systems (RSS) (2024)

    Ren, A.Z., Clark, J., Dixit, A., Itkina, M., Majumdar, A., Sadigh, D.: Explore un- til confident: Efficient exploration for embodied question answering. In: Robotics: Science and Systems (RSS) (2024)

    work page 2024

  26. [26]

    In: Proceedings of the 9th Conference on Robot Learning (CoRL) (2025)

    Saxena, S., Buchanan, B., Paxton, C., Liu, P., Chen, B., Vaskevicius, N., Palmieri, L., Francis, J., Kroemer, O.: GraphEQA: Using 3d semantic scene graphs for real- time embodied question answering. In: Proceedings of the 9th Conference on Robot Learning (CoRL) (2025)

    work page 2025

  27. [27]

    In: Proceedings of the Conference on Robot Learning (CoRL) (2023) Hypothesis Graph Refinement 17

    Shah,D.,Equi,M.,Osinski,B.,Xia,F.,Ichter,B.,Levine,S.:Navigationwithlarge language models: Semantic guesswork as a heuristic for planning. In: Proceedings of the Conference on Robot Learning (CoRL) (2023) Hypothesis Graph Refinement 17

    work page 2023

  28. [28]

    In: Findings of the Association for Computational Linguistics: ACL 2024

    Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L., Wang, Y.X., Yang, Y., Keutzer, K., Darrell, T.: Aligning large multimodal models with factually augmented RLHF. In: Findings of the Association for Computational Linguistics: ACL 2024. pp. 13088–13110. Association for Computational Linguistics, Bangkok, Thailand (Aug 2024)

    work page 2024

  29. [29]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Wald, J., Dhamo, H., Navab, N., Tombari, F.: Learning 3d semantic scene graphs from 3d indoor reconstructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3961–3970 (2020)

    work page 2020

  30. [30]

    arXiv preprint arXiv:2511.08935 (2025)

    Wang, N., Chen, W., Chen, L., Ji, H., Guo, Z., Zhang, X., Sun, H.: Expand your scope: Semantic cognition over potential-based exploration for embodied visual navigation. arXiv preprint arXiv:2511.08935 (2025)

    work page arXiv 2025

  31. [31]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Yang, Y., Yang, H., Zhou, J., Chen, P., Zhang, H., Du, Y., Gan, C.: 3D-Mem: 3d scene memory for embodied exploration and reasoning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17294–17303 (2025)

    work page 2025

  32. [32]

    arXiv preprint arXiv:2511.05894 (2025)

    Yu,F.,Deng,Q.,Tang,S.,Li,Y.,Cheng,L.:Open-world3dscenegraphgeneration for retrieval-augmented reasoning. arXiv preprint arXiv:2511.05894 (2025)

    work page arXiv 2025

  33. [33]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2025)

    Zemskova, T., Yudin, D.A.: 3DGraphLLM: Combining semantic graphs and large language models for 3d scene understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2025)

    work page 2025

  34. [34]

    arXiv preprint arXiv:2509.12618 (2025)

    Zhang, Z., Zhu, W., Pan, H., Wang, X., Xu, R., Sun, X., Zheng, F.: ActiveVLN: Towards active exploration via multi-turn RL in vision-and-language navigation. arXiv preprint arXiv:2509.12618 (2025)

    work page arXiv 2025

  35. [35]

    kitchen” with 0.75 confidence and expects objects {stove, refrigerator, sink}. Upon arrival, the agent observes a “bathroom

    Zhou, K., Zheng, K., Pryor, C., Shen, Y., Jin, H., Getoor, L., Wang, X.E.: Esc: Exploration with soft commonsense constraints for zero-shot object navigation. In: International Conference on Machine Learning. pp. 42829–42842. PMLR (2023) 18 P. Chen et al. A Detailed Comparison with Commonsense-Guided Exploration Table 7 compares HGR with recent commonsens...

    work page 2023

  36. [36]

    The VLM pre- dicts “kitchen” withρ= 0.82, creating hypothesis nodev hyp A with edge (vobs 1 , vhyp A ,0.82)

    The agent observes a frontierfA at the end of the hallway. The VLM pre- dicts “kitchen” withρ= 0.82, creating hypothesis nodev hyp A with edge (vobs 1 , vhyp A ,0.82)

    work page

  37. [37]

    stove” (ρ= 0.71),“refrigerator

    The VLM further predicts objects in the hypothesized kitchen: “stove” (ρ= 0.71),“refrigerator” (ρ= 0.68).Objecthypothesisnodesv obj A1 , vobj A2 arecreated with edges(v hyp A , vobj A1 ,0.71)and(v hyp A , vobj A2 ,0.68)

    work page

  38. [38]

    bedroom” withρ= 0.65, creatingv hyp B with edge(v obs 1 , vhyp B ,0.65)and child object “bed

    Another frontierfB is detected. The VLM predicts “bedroom” withρ= 0.65, creatingv hyp B with edge(v obs 1 , vhyp B ,0.65)and child object “bed” (ρ= 0.74)

    work page

  39. [39]

    kitchen” node decayed to 0.3 confidence still suggests nearby “stove

    The agent navigates tof A and observes a laundry room instead. Since ∆sem = 0.72>0.5, nodev hyp A is refuted. Cascade correction removesvhyp A , vobj A1 (stove), andv obj A2 (refrigerator)—three nodes total. The bedroom hy- pothesis branch (vhyp B and its children) is unaffected. C.3 Cascade Breadth Analysis In practice, dependency chains are shallow. Acr...

    work page