Pith. sign in

REVIEW 4 major objections 6 minor 43 references

Hyperbolic scene-graph memory replay cuts navigation forgetting to 28.5%.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 08:47 UTC pith:KMIMFGD7

load-bearing objection The scene-graph-plus-hyperbolic replay idea is new, but the Drop metric as defined doesn't measure retention, and the paper needs a corrected protocol before its claims can be trusted. the 4 major comments →

arxiv 2607.16267 v1 pith:KMIMFGD7 submitted 2026-07-04 cs.RO

HyperDCM: Dynamic Cluster Memory Replay in Hyperbolic Space for Continual Robotic Navigation Across Scenes

Zhengfei Lu , Jian Yang , Muyu Wang , Shaowen Chen , Jinpeng Mi , Ke Li , Xiong You , Qi Wu
show 2 more authors
Xuan Tang Xian Wei
This is my paper
classification cs.RO
keywords continual learningcatastrophic forgettingmemory replaydiffusion policyvisual navigationscene graphshyperbolic embeddingsPoincaré disk
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to show that the way a replay buffer is organized—not just how many old samples it keeps—determines how well a continually trained robot navigator holds onto earlier scenes. Its method, HyperDCM, turns each saved observation into a scene graph (subject–predicate–object triples), embeds that graph with a graph convolutional network, and projects the embedding into hyperbolic space, where hierarchical scene structure separates more cleanly. A dynamic clustering scheme keeps the buffer diverse by replacing each cluster's most central sample whenever a new, more distant sample arrives, and replay mixes samples across clusters with current training batches. In sequential training over five indoor and outdoor scenes, this reduces the success-rate drop from 43.3% (no replay) to 28.5%, and in a simulated closed-loop evaluation raises success from 39.2% to 56.0% while cutting collisions per meter from 0.37 to 0.17. The paper's point is that geometry-aware memory replay offers a practical route to adding scenes without retraining on all past data.

Core claim

The central claim is that replaying structured scene-graph samples from a hyperbolic-space memory buffer prevents catastrophic forgetting in diffusion-policy navigation better than uniform or Euclidean replay. The authors report that the full pipeline with Poincaré embeddings achieves a Drop of 28.5% on their five-scene sequential benchmark, versus 36.9% for uniform random replay, 35.2% for the same pipeline without scene graphs, and 34.8% for Euclidean-distance clustering; it also holds the highest success rate on intermediate scenes. In a closed-loop simulation, the same mechanism raises success rate from 39.2% to 56.0% and reduces collisions from 0.37 to 0.17 per meter relative to replay-

What carries the argument

The load-bearing component is the structure-sensitive memory update rule operating on clusters in the Poincaré disk. Each stored sample is represented as a scene graph, embedded to a 128-dimensional vector by a graph convolutional encoder that respects relation types, then mapped into hyperbolic space by a tangent-hyperbolic projection with fixed curvature. New samples are assigned to the nearest cluster center using Euclidean distance; each cluster's centroid is maintained as a Karcher mean, the Riemannian generalization of an average. When a cluster is full, the sample closest to the centroid is replaced by any incoming sample that is hyperbolically farther from the centroid—a rule that ke

Load-bearing premise

The whole advantage rests on the assumption that the (subject, predicate, object) triples produced by the vision-language model and the graph encoder yield embeddings whose hyperbolic projection separates navigation-relevant structure—if those triples are noisy, the projection is mis-specified (for example, curvature pushes samples outside the usable disk), or the replacement rule merely keeps outliers, the reported 28.5% drop advantage would collapse.

What would settle it

A single controlled comparison would settle the claim: train the same pipeline but scramble the predicate labels in the scene triples, keeping syntax and all hyperparameters fixed; if the success-rate drop stays near 28.5%, the scene-graph structure is not carrying the benefit. A second check is to set the curvature parameter c to zero so the Poincaré distance degenerates toward Euclidean distance; if the drop does not rise toward the Euclidean variant's 34.8%, the reported geometric effect is not real.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the claim is right, a fixed-size memory of 1,280 samples can support continual navigation across scenes, cutting forgetting to a 28.5% drop without joint retraining.
  • The ordering across ablations—uniform replay, visual-feature replay, Euclidean structured replay, hyperbolic structured replay—shows that structural representation and hyperbolic geometry each contribute independently to stability.
  • The approach transfers from open-loop waypoint prediction to closed-loop simulation, improving success rate and path efficiency while reducing collision frequency.
  • Per-stage training cost stays nearly constant as scenes accumulate, unlike joint retraining whose cost grows linearly, so the method adapts a new scene in the same training budget rather than revisiting all past data.
  • The conclusion generalizes beyond navigation, proposing structural world modeling plus hyperbolic memory as a general anti-forgetting recipe for robotic learning.

Where Pith is reading between the lines

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

  • A natural test the paper's comparisons do not fully settle is whether the benefit comes from the scene-graph content or merely from the clustering rule; replacing predicates with random labels under the same pipeline would isolate the semantic contribution.
  • Because the graph encoder is fixed offline and its training is not described, the method's sensitivity to that encoder is unknown; retraining it on different scene-graph supervision could shift the reported gaps.
  • The success metric (over 90% waypoint alignment) is coarse; the Drop ordering might change at stricter thresholds, so the retention advantage deserves scrutiny at multiple alignment thresholds.
  • The replacement rule favors samples at the periphery of each cluster, which could select outlier views; a variant that maximizes minimum pairwise separation within the cluster, rather than distance to centroid, may be more stable.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes HyperDCM, a continual-learning replay mechanism for diffusion-policy visual navigation. It extracts semantic (subject, predicate, object) triples from RGB images using Qwen-VL/Qwen3-7B, embeds them as scene graphs with an R-GCN, projects the embeddings into a Poincaré disk, and maintains a multi-cluster memory with a centroid-based replacement rule. The method is evaluated on SACSoN and Recon with NoMaD as base policy, plus a Habitat simulation, reporting a Drop of 28.5% for HyperDCM-Hyp versus 34.8% for Euclidean replay and improved SR/SPL/collision in Habitat. The claim is that structure-aware hyperbolic replay mitigates catastrophic forgetting better than uniform or Euclidean replay.

Significance. If the retention claim were properly established, the paper would be a useful contribution: it combines scene-graph structure, hyperbolic geometry, and replay in a relatively underexplored setting, and the HyperDCM-Hyp versus HyperDCM-Euc ablation is a clean way to isolate the geometric component. The paper, however, does not currently provide a valid measure of forgetting: the reported Drop metric compares success on different scenes at different checkpoints, and the hyperbolic projection is mis-specified. The absence of R-GCN training details and key hyperparameter values further weakens reproducibility. The core claim therefore is not yet supported by the evidence.

major comments (4)
  1. [Section IV-C, Table I] The Drop metric is defined as SR(1~A) − SR(1~E). Given the protocol, 1~X is the checkpoint after sequential training through Scene X, and Table I reports the SR for the corresponding current scene. Under this definition, Drop compares SR on Scene A at checkpoint 1~A with SR on Scene E at checkpoint 1~E. This is a cross-scene, cross-checkpoint difference; it does not quantify how much performance on earlier scenes (e.g., Scene A) degrades after training on later scenes. The paper's central claim of 'superior retention of past navigation capabilities' (Abstract, Section V) requires per-scene retention curves, e.g., SR_SceneA at 1~A versus 1~E. Such numbers are never reported. Consequently, the headline comparison (28.5% vs 34.8% Drop) does not support the retention claim; the higher 1~E SR of HyperDCM-Hyp could reflect better adaptation to the final scene rather than preserved memory. If t
  2. [Section III-B, Eq. (3)] The proposed Poincaré projection is not the standard exponential map. With the usual convention for curvature c>0, exp_0(v)=tanh(√c||v||) v/(√c||v||), whose norm is bounded by 1/√c. The formula in Eq. (3), φ(z_i)=z_i√c(||z_i||+ε) tanh(√c||z_i||), has norm √c(||z_i||+ε) tanh(√c||z_i||), which can exceed 1/√c whenever ||z_i||+ε > 1/√c. Points can therefore leave the Poincaré disk, making the distance in Eq. (5) undefined and the Karcher mean computation ill-posed. Unless the formula is corrected (or c and ε are specified so that the image remains in the disk), the hyperbolic geometry component is not well defined.
  3. [Section III-A/IV-A, Eq. (1)] The R-GCN is 'used solely for offline embedding generation' and 'fixed after preprocessing,' but the manuscript gives no information about how it is trained: no relation vocabulary, supervision signal, training set, or loss. Since all downstream memory operations act on R-GCN embeddings, this is a missing load-bearing component. In addition, the curvature c, clustering threshold τ, cluster count C, capacity M, replay batch size n, and success threshold δ are never specified; c and τ are only described as 'determined beforehand and kept unchanged.' Without these values and a sensitivity analysis, the reported improvements are not reproducible and could be tuned.
  4. [Sections IV-E/IV-F, Tables I-II and Fig. 5] Neither table provides evidence of per-scene retention. Table II reports aggregate SR/SPL/collision in Habitat, which can improve due to better current-scene adaptation; Fig. 5's 'GC cosine similarity' is not formally defined and is not linked to the five-scene protocol. No standard deviation, confidence interval, or number of seeds is reported anywhere. As a result, even the secondary claims about 'intermediate scenes' are not statistically supported.
minor comments (6)
  1. [Section III-C] Eq. (6) repeats the Karcher mean definition of Eq. (4); consolidate and define the domain D^n.
  2. [Fig. 2] UMAP axes are unlabeled; the claim of 'better separation' is not quantified (e.g., silhouette score or average hyperbolic distance between clusters).
  3. [Fig. 5] 'GC cosine similarity' should be defined; currently it is not possible to interpret the drop/improvement values without a formal definition.
  4. [Table I] Drop values are computed from rounded percentages; report exact values or at least one decimal place to avoid spurious precision in the differences.
  5. [Section IV-D] Some baselines are underspecified: EWC's Fisher estimate, LwF's distillation loss weight, and TinyER's buffer management are not detailed; a short paragraph or reference to implementation details is needed.
  6. [General] No code or model weights are provided; for reproducibility, please release them or provide a detailed implementation appendix. Also clarify whether the Habitat scenes are a separate benchmark or a re-simulation of SACSoN/Recon.

Circularity Check

1 steps flagged

The headline retention metric (Drop) is defined as a current-scene SR gap and relabeled as catastrophic forgetting, so the central retention claim reduces to the paper's own definition.

specific steps
  1. renaming known result [Section IV-C, Evaluation Metrics and Protocol (Drop definition; used in Table I)]
    "We additionally report the Drop metric, defined as the SR difference between the first and last scenes in the sequence (1∼Avs.1∼E). A larger Drop indicates more severe catastrophic forgetting, providing a direct measure of knowledge retention across continual training: Drop=SR 1∼A −SR 1∼E. ... Evaluation is conducted using the checkpoint obtained at the end of the corresponding training sequence."

    By the paper's own protocol, 1∼A is the checkpoint after training only Scene A, evaluated on Scene A, while 1∼E is the checkpoint after training Scenes 1–E, evaluated on Scene E. The Drop therefore compares two different scenes at two different checkpoints; it never re-tests an earlier scene after later training (e.g., Scene A after B–E). The headline '28.5% Drop' is exactly this defined difference read from Table I, relabeled as 'knowledge retention.' The claimed retention advantage is an artifact of the chosen definition, not an independent measurement of forgetting.

full rationale

The paper is an empirical system paper with no formal derivation; the geometric comparison HyperDCM-Hyp vs HyperDCM-Euc is a clean ablation and would be informative if Drop actually measured retention. However, the paper defines Drop = SR(1∼A) − SR(1∼E), where 1∼A and 1∼E are different scenes at different checkpoints. This never measures how much an earlier scene's success degrades after later training, so the reported 28.5% vs 34.8% Drop is a current-scene success gap renamed as catastrophic forgetting. The central retention claim therefore reduces to the paper's own definitional choice. The self-citations ([6], [23]) are not load-bearing, and the predetermined c/τ values are a reporting concern rather than circularity. Because the headline retention metric is constructed rather than measured, the paper receives 6; the Habitat results (Table II) provide independent evidence of adaptation gains but do not repair the retention metric.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim rests on several undisclosed constants (c, τ, C, M, n, δ) and on unvalidated assumptions about LLM scene-graph quality and R-GCN embedding hierarchy. The projection formula may be mathematically unsound, which would invalidate the hyperbolic geometry premise. No new physical entities are introduced; 'Hyperbolic Dynamic Cluster Memory' is a software mechanism, and 'structural richness' is an informal quantity without a formal definition.

free parameters (6)
  • curvature c
    Used in Poincaré map and hyperbolic distance; 'fixed curvature c across all scenes, determined beforehand and kept unchanged during final evaluation' (Sec III-B). Value and selection procedure are not reported.
  • clustering threshold τ
    Determines when a new cluster is created; 'fixed clustering threshold τ across all scenes' (Sec III-C); value not reported; can change memory composition and hence replay quality.
  • cluster count C and per-cluster capacity M
    Sec III-C defines memory as C clusters of capacity M; Table I says replay capacity is 1280 samples, but neither C nor M is specified.
  • replay batch size n
    Sec III-D: 'uniformly sample n instances from the C memory clusters'; n not reported; replay batch composition directly affects forgetting.
  • success threshold δ
    Appears in Eq. (1) as the success threshold d(s_i,g_i) ≤ δ, but never set; then the prose replaces the criterion with the 90%-waypoint rule.
  • R-GCN weights
    Sec III-A: 'weights follow the standard R-GCN formulation and are fixed after preprocessing' — the training that produces these weights is not described; effectively a fitted component of the pipeline.
axioms (5)
  • domain assumption Qwen-VL and Qwen3-7B produce accurate, navigation-relevant scene triples from RGB images
    Stated in Sec III-A ('These large models provide reliable relational extraction'); no validation or error analysis is provided for the navigation setting.
  • domain assumption R-GCN embeddings, trained offline, encode hierarchy useful for navigation
    Sec III-A says weights are 'fixed after preprocessing' but training data/labels are not described; the downstream usefulness is assumed.
  • ad hoc to paper Samples with richer relational structure lie farther from the origin in the Poincaré disk
    Sec III-B asserts this ('Samples with richer relational structures are naturally positioned farther from the origin'), and it is the basis of the replacement rule; no proof or measurement is given.
  • ad hoc to paper Replacing the centroid-closest sample with a farther new sample preserves memory diversity and reduces forgetting
    Sec III-C formalizes the rule; it is a design choice with no analysis showing outliers are the most useful replay samples.
  • standard math Standard R-GCN and Poincaré-disk formulas are valid as used
    Sec III-A/B invokes standard formulations, but the projection formula appears inconsistent with the standard exponential map and may leave the disk; the distance formula is valid only inside the disk.

pith-pipeline@v1.3.0-alltime-deepseek · 10056 in / 19508 out tokens · 204579 ms · 2026-08-02T08:47:37.914550+00:00 · methodology

0 comments
read the original abstract

Continual learning in visual navigation remains challenging due to catastrophic forgetting and the difficulties associated with adapting to diverse and evolving environments. To address these issues, we propose Hyperbolic Dynamic Cluster Memory (HyperDCM), a structure-aware memory mechanism that enhances diffusion policy-based navigation through scene graph modeling and principled memory replay. HyperDCM extracts semantic scene triples from RGB observations using large vision-language models, encodes them into scene graph embeddings via a Relational Graph Convolutional Network (R-GCN), and projects the embeddings into hyperbolic space to enhance structural separability and retention in continual navigation. A dynamic clustering and structure-sensitive update strategy selects representative samples for memory replay, thereby preserving knowledge diversity and mitigating catastrophic forgetting. Experiments on multi-scene indoor and outdoor datasets demonstrate that HyperDCM achieves superior retention of past navigation capabilities and improved generalization compared to representative continual learning baselines adapted to diffusion policy navigation.

Figures

Figures reproduced from arXiv: 2607.16267 by Jian Yang, Jinpeng Mi, Ke Li, Muyu Wang, Qi Wu, Shaowen Chen, Xian Wei, Xiong You, Xuan Tang, Zhengfei Lu.

Figure 1
Figure 1. Figure 1: Overview of the proposed HyperDCM mechanism. Navigation data from robot trajectories is collected, where RGB observations are processed by vision–language models to generate semantic descriptions, which are parsed into (Subject, Predicate, Object) triples. These triples are used to construct scene graphs, encoded via an R-GCN, and projected into hyperbolic space to capture hierarchical relations. During di… view at source ↗
Figure 2
Figure 2. Figure 2: UMAP visualization comparing the Euclidean and Poincar [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: We divide memory into C clusters, each with capacity M. Here, C denotes the number of memory clusters and M is the per-cluster capacity; dE(·, ·) and dH(·, ·) denote Euclidean and hyperbolic geodesic distances, respectively. For each new scene embedding ϕ(zi): Compute the Euclidean distance dE to the center of all clusters {ck} C k=1: min k∈{1,...,C} dE(ϕ(zi), ck). If the minimum distance exceeds a thresho… view at source ↗
Figure 3
Figure 3. Figure 3: Dynamic cluster memory management in hyperbolic space. Each new sample computes its Euclidean distance to all cluster centroids, being [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of predicted waypoint trajectories across four representative scenes. Left: continual NoMaD baseline. Right: HyperDCM [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Trajectory-level forgetting analysis on goal-conditioned (GC) met [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison between joint retraining and HyperDCM under sequen [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

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

43 extracted references · 6 linked inside Pith

  1. [1]

    Safe visual navigation via deep learning and novelty detection,

    C. Richter and N. Roy, “Safe visual navigation via deep learning and novelty detection,” 2017

    2017

  2. [2]

    Cog- nitive mapping and planning for visual navigation,

    S. Gupta, J. Davidson, S. Levine, R. Sukthankar, and J. Malik, “Cog- nitive mapping and planning for visual navigation,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2616–2625

    2017

  3. [3]

    Visual navigation with spatial attention,

    B. Mayo, T. Hazan, and A. Tal, “Visual navigation with spatial attention,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 16 898–16 907

    2021

  4. [4]

    Vint: A foundation model for visual navigation,

    D. Shah, A. Sridhar, N. Dashora, K. Stachowicz, K. Black, N. Hirose, and S. Levine, “Vint: A foundation model for visual navigation,” in Proceedings of The 7th Conference on Robot Learning, vol. 229. PMLR, 2023, pp. 711–733

    2023

  5. [5]

    Ok-robot: What really matters in integrating open-knowledge models for robotics,

    P. Liu, Y . Orru, J. Vakil, C. Paxton, N. M. M. Shafiullah, and L. Pinto, “Ok-robot: What really matters in integrating open-knowledge models for robotics,”arXiv preprint arXiv:2401.12202, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    Dual-bev nav: Dual-layer bev-based heuristic path planning for robotic navigation in unstructured outdoor environments,

    J. Zhang, H. Dong, J. Yang, J. Liu, S. Huang, K. Li, X. Tang, X. Wei, and X. You, “Dual-bev nav: Dual-layer bev-based heuristic path planning for robotic navigation in unstructured outdoor environments,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 8872–8879

    2025

  7. [7]

    Diffusion policy: Visuomotor policy learning via ac- tion diffusion,

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song, “Diffusion policy: Visuomotor policy learning via ac- tion diffusion,”The International Journal of Robotics Research, p. 02783649241273668, 2023

    2023

  8. [8]

    Nomad: Goal masked diffusion policies for navigation and exploration,

    A. Sridhar, D. Shah, C. Glossop, and S. Levine, “Nomad: Goal masked diffusion policies for navigation and exploration,” in2024 IEEE International Conference on Robotics and Automation. IEEE, 2024, pp. 63–70

    2024

  9. [9]

    Open-world continual learning: Unifying novelty detection and continual learning,

    G. Kim, C. Xiao, T. Konishi, Z. Ke, and B. Liu, “Open-world continual learning: Unifying novelty detection and continual learning,”Artificial Intelligence, vol. 338, p. 104237, 2025

    2025

  10. [10]

    Online task-free continual generative and discriminative learning via dynamic cluster memory,

    F. Ye and A. G. Bors, “Online task-free continual generative and discriminative learning via dynamic cluster memory,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, 2024, pp. 26 202–26 212

    2024

  11. [11]

    Domain incremental lifelong learning in an open world,

    Y . Dai, H. Lang, Y . Zheng, B. Yu, F. Huang, and Y . Li, “Domain incremental lifelong learning in an open world,” inFindings of the Association for Computational Linguistics: ACL 2023, 2023, pp. 5844–5865

    2023

  12. [12]

    Ai autonomy: Self-initiated open-world continual learning and adaptation,

    B. Liu, S. Mazumder, E. Robertson, and S. Grigsby, “Ai autonomy: Self-initiated open-world continual learning and adaptation,”AI Mag- azine, vol. 44, no. 2, pp. 185–199, 2023

    2023

  13. [13]

    Target search and navigation in heterogeneous robot systems with deep reinforcement learning,

    Y . Chen and J. Xiao, “Target search and navigation in heterogeneous robot systems with deep reinforcement learning,”Machine Intelligence Research, vol. 22, no. 1, pp. 79–90, 2025

    2025

  14. [14]

    Navformer: A transformer ar- chitecture for robot target-driven navigation in unknown and dynamic environments,

    H. Wang, A. H. Tan, and G. Nejat, “Navformer: A transformer ar- chitecture for robot target-driven navigation in unknown and dynamic environments,”IEEE Robotics and Automation Letters, vol. 9, no. 8, pp. 6808–6815, 2024

    2024

  15. [15]

    Brain-inspired replay for continual learning with artificial neural networks,

    G. M. Van de Ven, H. T. Siegelmann, and A. S. Tolias, “Brain-inspired replay for continual learning with artificial neural networks,”Nature communications, vol. 11, no. 1, p. 4069, 2020

    2020

  16. [16]

    Scene graph generation from objects, phrases and region captions,

    Y . Li, W. Ouyang, B. Zhou, K. Wang, and X. Wang, “Scene graph generation from objects, phrases and region captions,” inProceedings of the IEEE international conference on computer vision, 2017, pp. 1261–1270

    2017

  17. [17]

    Hyperbolic deep neural networks: A survey,

    W. Peng, T. Varanka, A. Mostafa, H. Shi, and G. Zhao, “Hyperbolic deep neural networks: A survey,”IEEE Transactions on pattern analysis and machine intelligence, vol. 44, no. 12, pp. 10 023–10 044, 2021

    2021

  18. [18]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,”Advances in neural information processing systems, vol. 33, pp. 6840–6851, 2020

    2020

  19. [19]

    Diffusion models in vision: A survey,

    F.-A. Croitoru, V . Hondru, R. T. Ionescu, and M. Shah, “Diffusion models in vision: A survey,”IEEE transactions on pattern analysis and machine intelligence, vol. 45, no. 9, pp. 10 850–10 869, 2023

    2023

  20. [20]

    Versatile navigation under partial observability via value-guided diffusion policy,

    G. Zhang, H. Tang, and Y . Yan, “Versatile navigation under partial observability via value-guided diffusion policy,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 17 943–17 951

    2024

  21. [21]

    Navidiffusor: Cost-guided diffusion model for visual navigation,

    Y . Zeng, H. Ren, S. Wang, J. Huang, and H. Cheng, “Navidiffusor: Cost-guided diffusion model for visual navigation,”arXiv preprint arXiv:2504.10003, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    Dare: Diffusion policy for autonomous robot exploration,

    Y . Cao, J. Lew, J. Liang, J. Cheng, and G. Sartoretti, “Dare: Diffusion policy for autonomous robot exploration,” in2025 IEEE International Conference on Robotics and Automation. IEEE, 2025, pp. 11 987– 11 993

    2025

  23. [23]

    Kiterunner: Language-driven cooperative local-global navigation policy with uav mapping in outdoor environ- ments,

    S. Huang, C. Shi, J. Yang, H. Dong, J. Mi, K. Li, J. Zhang, M. Ding, P. Liang, X. Youet al., “Kiterunner: Language-driven cooperative local-global navigation policy with uav mapping in outdoor environ- ments,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2025, pp. 18 512–18 519

    2025

  24. [24]

    Ldp: A local diffusion planner for efficient robot navigation and collision avoidance,

    W. Yu, J. Peng, H. Yang, J. Zhang, Y . Duan, J. Ji, and Y . Zhang, “Ldp: A local diffusion planner for efficient robot navigation and collision avoidance,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 2024, pp. 5466–5472

    2024

  25. [25]

    A comprehensive survey of continual learning: Theory, method and application,

    L. Wang, X. Zhang, H. Su, and J. Zhu, “A comprehensive survey of continual learning: Theory, method and application,”IEEE transac- tions on pattern analysis and machine intelligence, vol. 46, no. 8, pp. 5362–5383, 2024

    2024

  26. [26]

    Overcoming catastrophic forgetting in neural networks,

    J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska et al., “Overcoming catastrophic forgetting in neural networks,”Pro- ceedings of the national academy of sciences, vol. 114, no. 13, pp. 3521–3526, 2017

    2017

  27. [27]

    Continual learning through synap- tic intelligence,

    F. Zenke, B. Poole, and S. Ganguli, “Continual learning through synap- tic intelligence,” inInternational conference on machine learning. PMLR, 2017, pp. 3987–3995

    2017

  28. [28]

    Learning without forgetting,

    Z. Li and D. Hoiem, “Learning without forgetting,”IEEE transactions on pattern analysis and machine intelligence, vol. 40, no. 12, pp. 2935–2947, 2017

    2017

  29. [29]

    icarl: Incremental classifier and representation learning,

    S.-A. Rebuffi, A. Kolesnikov, G. Sperl, and C. H. Lampert, “icarl: Incremental classifier and representation learning,” inProceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2017, pp. 2001–2010

    2017

  30. [30]

    On tiny episodic memories in continual learning,

    A. Chaudhry, M. Rohrbach, M. Elhoseiny, T. Ajanthan, P. K. Dokania, P. H. Torr, and M. Ranzato, “On tiny episodic memories in continual learning,”arXiv preprint arXiv:1902.10486, 2019

    Pith/arXiv arXiv 1902

  31. [31]

    Packnet: Adding multiple tasks to a single network by iterative pruning,

    A. Mallya and S. Lazebnik, “Packnet: Adding multiple tasks to a single network by iterative pruning,” inProceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2018, pp. 7765–7773

    2018

  32. [32]

    Overcoming catastrophic forgetting with hard attention to the task,

    J. Serra, D. Suris, M. Miron, and A. Karatzoglou, “Overcoming catastrophic forgetting with hard attention to the task,” inInternational conference on machine learning. PMLR, 2018, pp. 4548–4557

    2018

  33. [33]

    Stable continual reinforcement learning via diffusion-based trajectory replay,

    F. Chen, F. Han, C. Guan, L. Yuan, Z. Zhang, Y . Yu, and Z. Zhang, “Stable continual reinforcement learning via diffusion-based trajectory replay,”arXiv preprint arXiv:2411.10809, 2024

    Pith/arXiv arXiv 2024

  34. [34]

    Continual diffusion with stamina: Stack-and-mask incremental adapters,

    J. S. Smith, Y .-C. Hsu, Z. Kira, Y . Shen, and H. Jin, “Continual diffusion with stamina: Stack-and-mask incremental adapters,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 1744–1754

    2024

  35. [35]

    Continual diffusion: Continual customization of text-to-image diffusion with c-lora,

    J. S. Smith, Y .-C. Hsu, L. Zhang, T. Hua, Z. Kira, Y . Shen, and H. Jin, “Continual diffusion: Continual customization of text-to-image diffusion with c-lora,”arXiv preprint arXiv:2304.06027, 2023

    Pith/arXiv arXiv 2023

  36. [36]

    Adaptive visual scene un- derstanding: incremental scene graph generation,

    N. Khandelwal, X. Liu, and M. Zhang, “Adaptive visual scene un- derstanding: incremental scene graph generation,” inProceedings of the 38th International Conference on Neural Information Processing Systems, 2024, pp. 136 446–136 478

    2024

  37. [37]

    Graph r-cnn for scene graph generation,

    J. Yang, J. Lu, S. Lee, D. Batra, and D. Parikh, “Graph r-cnn for scene graph generation,” inProceedings of the European conference on computer vision, 2018, pp. 670–685

    2018

  38. [38]

    Hyperbolic neural net- works,

    O. Ganea, G. B ´ecigneul, and T. Hofmann, “Hyperbolic neural net- works,”Advances in neural information processing systems, vol. 31, 2018

    2018

  39. [39]

    Poincar ´e embeddings for learning hierar- chical representations,

    M. Nickel and D. Kiela, “Poincar ´e embeddings for learning hierar- chical representations,”Advances in neural information processing systems, vol. 30, 2017

    2017

  40. [40]

    Sacson: Scalable autonomous control for social navigation,

    N. Hirose, D. Shah, A. Sridhar, and S. Levine, “Sacson: Scalable autonomous control for social navigation,”IEEE Robotics and Au- tomation Letters, vol. 9, no. 1, pp. 49–56, 2023

    2023

  41. [41]

    Rapid ex- ploration for open-world navigation with latent goal models,

    D. Shah, B. Eysenbach, N. Rhinehart, and S. Levine, “Rapid ex- ploration for open-world navigation with latent goal models,” in Conference on Robot Learning. PMLR, 2022, pp. 674–684

    2022

  42. [42]

    Matterport3d: Learning from rgb-d data in indoor environments,

    A. Chang, A. Dai, T. Funkhouser, M. Halber, M. Niessner, M. Savva, S. Song, A. Zeng, and Y . Zhang, “Matterport3d: Learning from rgb-d data in indoor environments,”arXiv preprint arXiv:1709.06158, 2017

    Pith/arXiv arXiv 2017

  43. [43]

    Habitat: A platform for embodied ai research,

    M. Savva, A. Kadian, O. Maksymets, Y . Zhao, E. Wijmans, B. Jain, J. Straub, J. Liu, V . Koltun, J. Maliket al., “Habitat: A platform for embodied ai research,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 9339–9347

    2019