Pith. sign in

REVIEW 4 major objections 6 minor 42 references

TravKAN: Fast and Interpretable Nonlinear Traversability Analysis with Kolmogorov-Arnold Networks

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper shows that a shallow Kolmogorov-Arnold network, trained on handcrafted LiDAR features including new reflectivity descriptors, can match the traversability-ranking performance of XGBoost while allowing its decision function to be

desk verdict The empirical claim holds up—a shallow KAN is competitive with XGBoost on three traversability benchmarks—but the interpretability story is unsupported as written because TravKAN-Lite is a poor surrogate on two datasets and the symbolic expression is not in the paper. read the letter →

arxiv 2608.02320 v1 pith:PVILWLK5 submitted 2026-08-03 cs.RO cs.CV

Daniel Fusaro , Simone Mosco , Wanmeng Li , Alberto Pretto This is my paper · ORCID
classification cs.ROcs.CV
keywords traversabilityanalysisKolmogorov-ArnoldNetworksLiDARreflectivityfeaturesinterpretablemachinelearningoff-roadroboticspolargridsymbolicextraction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

TravKAN argues that Kolmogorov-Arnold Networks, which learn univariate functions on edges instead of using fixed activations, can make LiDAR-based traversability classification both accurate and interpretable. Using handcrafted geometric, remission, and newly introduced reflectivity features aggregated on polar grid cells, a single-hidden-layer KAN matches or approaches the ranking performance of XGBoost and beats deep MLP and Transformer baselines across three real-world datasets. After training, the learned network can be collapsed into a closed-form analytic expression, TravKAN-Lite, that runs in 0.16 ms per frame on CPU and exposes which terrain statistics dominate the decision. If right, the paper offers a practical route to transparent, real-time traversability reasoning for safety-critical robot navigation.

What carries the argument

Kolmogorov-Arnold Networks (KANs): each edge of the network is parameterized as a learnable univariate B-spline function, so a layer maps inputs via sums of edge functions rather than fixed nonlinearities followed by affine maps. This edge-based spline parameterization is what enables both universal approximation through layer composition and symbolic extraction of each learned function after training. The paper couples this network with a polar-grid cell representation and handcrafted features—geometric, remission, and a new set of reflectivity descriptors—that feed a [d, 24, 2] network.

What would settle it

Train the same pipeline on a dataset where terrain patches are labeled by physical outcomes (for example, measured robot slippage or rollover) rather than semantic class, and check whether TravKAN-Lite's symbolic expression still achieves competitive AUC; if its ranking performance degrades substantially, the semantic-label proxy rather than the KAN machinery is carrying the result.

Watch

Extended reading notes

Core claim

The central claim is that a single-hidden-layer KAN with 24 hidden units and B-spline parameterized edge functions is sufficient to approximate the traversability decision boundary from handcrafted LiDAR features as well as a 3000-tree XGBoost ensemble—and better than MLPs and Transformers on ranking metrics—while its learned edge functions can be symbolically extracted into a compact expression composed of linear, polynomial, exponential, trigonometric, and rational terms. The paper also introduces reflectivity-based handcrafted descriptors and shows a small but consistent improvement on the off-road EastPark dataset. The authors state that the resulting symbolic model, TravKAN-Lite, reveal

Load-bearing premise

The ground-truth traversability labels are derived by mapping semantic classes to binary labels, so the evaluation and the extracted symbolic rule describe agreement with semantic categories, not physical traversability for any specific robot with its own size, weight, and traction capabilities.

Editorial extensions

If this is right

  • A shallow, interpretable network can replace deep models for tabular traversability classification without loss of ranking accuracy, shrinking model size from megabytes to tens of kilobytes.
  • The symbolic expression TravKAN-Lite provides an analytic decision rule with effectively no memory footprint and fast enough inference for real-time on-device navigation.
  • Reflectivity, a LiDAR channel usually ignored in handcrafted descriptors, carries complementary material information that improves discrimination between terrains with similar geometry.
  • The extracted symbolic form makes feature attribution explicit, showing height statistics dominate and surface normals or unevenness act as local corrections, which can guide future feature design.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same KAN-based symbolic extraction could extend to other geometric perception tasks, such as terrain roughness estimation, soil bearing capacity, or planning cost learning, where a closed-form cost is desirable.
  • If the symbolic expression learned on EastPark transfers to a different LiDAR sensor or a different robot platform without retraining, that would indicate the model captures material properties rather than dataset-specific statistics; this is a testable extension the authors leave implicit.
  • The interpretability claim depends on how faithfully TravKAN-Lite approximates the trained TravKAN; measuring the fidelity gap across datasets would quantify how much accuracy transparency costs.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes TravKAN, a shallow Kolmogorov-Arnold Network trained on handcrafted LiDAR features (geometric, remission, and a newly introduced reflectivity group) for cell-wise binary traversability classification. The method is evaluated on SemanticKITTI, nuScenes, and EastPark against MLP, MLP++, TransformerEncoder, P-SVM, and XGBoost. Tables I–III show that TravKAN is broadly competitive with XGBoost in ROC AUC and PR AUC, sometimes slightly better, while using fewer parameters. The paper also presents TravKAN-Lite, a claimed symbolic extraction of the trained model, and reports ablations on architecture and reflectivity features. The central claims are that TravKAN is fast, interpretable, and competitive with strong baselines.

Significance. If the core empirical result holds, the paper demonstrates a useful point: a compact KAN can match gradient-boosted trees on tabular traversability features with a fraction of the parameters. The use of public datasets and the provision of a code link are strengths, as is the inclusion of multiple baselines and ablations. However, the paper's significance as an interpretability contribution is not currently supported: the symbolic expression is not included in the manuscript, and the reported TravKAN-Lite results show large fidelity losses on two of three datasets. The 'fast' claim also conflicts with the reported runtime of the full KAN model. These issues would need to be resolved before the paper's advertised contributions are established.

major comments (4)
  1. [V, Eq. (7); Tables II–III] The interpretability claim is load-bearing and currently unsupported. Eq. (7) is a template (L+Q+E+S+R+C+b), not an actual analytic expression; the full expression is deferred to a GitHub repository, and no symbolic-extraction or simplification procedure is described. More importantly, the reported fidelity of TravKAN-Lite is poor: PR AUC drops from 97.1 to 69.8 on SemanticKITTI (Table II) and from 98.9 to 91.3 on nuScenes (Table III), with balanced accuracy dropping 7.2 and 6.3 points, respectively. A 27-point PR-AUC drop means the symbolic form does not preserve the trained model's ranking behavior. The text in Section IV-E calling Lite 'competitive' is not supported. To substantiate the interpretability contribution, the paper should include the actual expression, describe the extraction method, and report a fidelity measure (e.g., agreement with TravKAN predictions or decision values
  2. [Abstract, Fig. 1, Table IV] The 'fast' claim is contradicted by the paper's own runtime table. Table IV reports TravKAN at 71 ms/frame CPU, versus 0.9 ms for MLP, 1.4 ms for MLP++, 5.3 ms for TransformerEncoder, 25 ms for XGBoost, and 50 ms for P-SVM. Thus TravKAN itself is slower than all baselines except P-SVM; only Lite-TravKAN (0.16 ms) is fastest. Fig. 1's caption claims 'extremely reduced runtime and O(1) computational complexity,' which is not accurate for the KAN model. Please revise the abstract and Fig. 1 to attribute speed to TravKAN-Lite only, and either report per-cell latency or explain why the full model's runtime is still acceptable for the target deployment setting.
  3. [Tables I–III] All performance numbers are single-run point estimates with no repeated seeds, error bars, or statistical tests. Several headline comparisons differ by less than one percentage point (e.g., EastPark balanced accuracy 87.6–87.9 across methods; nuScenes ROC AUC 99.1 for TravKAN and XGBoost), so the claims of 'outperforming' or 'approaching' are not robust to run-to-run variance. The paper should report mean ± std across multiple seeds and, ideally, paired significance tests for the key TravKAN-versus-XGBoost and TravKAN-versus-MLP++ comparisons.
  4. [IV-A, VI] The ground-truth labels are derived from semantic annotations by mapping drivable/flat classes to positive and obstacles/vegetation/vehicles/vertical structures to negative. This is a proxy for physical traversability and ignores platform-specific size, weight, and traction. Since all evaluations and the interpretability narrative rest on this label proxy, the paper should state this limitation explicitly and, if possible, validate on a robot-specific label set or report per-class results. In addition, the reflectivity features are evaluated only on EastPark (Table VI and Section IV-C), while the abstract claims reflectivity is 'systematically exploited.' The abstract and conclusions should be narrowed to reflect that the reflectivity contribution is demonstrated on one dataset.
minor comments (6)
  1. [IV-C] The sentence 'extended with reflectivity-based descriptors ... only for the EastPark dataset' is ambiguous. Clarify whether SemanticKITTI and nuScenes lack a reflectivity channel or whether the descriptors were simply not computed there.
  2. [Table IV] For P-SVM, #Params is reported as 30K while model size is 25 MB; this is inconsistent unless support vectors or kernel matrices are counted. Please explain what is included.
  3. [V] The function symbols L, Q, E, S, R, and C in Eq. (7) are introduced but never defined precisely. Even if the full expression is in a repository, the paper should give at least one concrete example term or the exact functional form used for each category.
  4. [Throughout] The naming is inconsistent: 'TravKan', 'TravKAN', 'Lite-TravKAN', and 'TravKan-Lite' are used interchangeably. Standardize to one name.
  5. [III-D] The statement that B-spline universality plus layer stacking 'yields universal approximation' is imprecise; the Kolmogorov-Arnold theorem places constraints on the number of layers and width. This is a theoretical aside, but it should be phrased carefully.
  6. [References] Reference [7] for EastPark is only a URL with no author/title/venue. Provide a full citation or indicate it is a dataset release.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: TravKAN is an empirical ML pipeline; no prediction reduces to a fitted parameter or self-citation.

full rationale

The paper's derivation chain is empirical rather than analytical: it extracts handcrafted LiDAR features, trains a shallow KAN, and evaluates on held-out test sets. The symbolic expression (Eq. 7) is a post-training extraction, not a prediction that is defined by its own inputs; TravKAN-Lite is evaluated like any compressed model. The main self-citation is to the authors' prior P-SVM2 work [11] for feature descriptors and the ground-truth labeling protocol. Although this prior work is load-bearing in the sense that the feature set is taken from it, it is an externally published, code-reproduced pipeline, and the present paper compares all methods under the same features and labels. There is no argument that reduces the central performance claim to the correctness of [11], and no 'uniqueness' or 'ansatz' is smuggled in via self-citation. The reflectivity features are empirically ablated, not derived from the target metric. The poor fidelity of TravKAN-Lite on SemanticKITTI and nuScenes is a support weakness for the interpretability claim, but it is not a circularity: the symbolic approximation is not treated as an independent prediction equal to its inputs. Overall, the central claim has independent empirical content, so the circularity score is low.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper's central claim is empirical, so the ledger captures the modeling choices and domain assumptions rather than derived free parameters. The learned KAN weights (spline coefficients, affine scalars) are fitted to data as part of training, but they are the model itself, not ad hoc structural parameters; the listed items are the hand-chosen hyperparameters and domain-level assumptions that the results depend on.

free parameters (4)
  • KAN hidden dimension = 24
    Chosen based on validation performance; affects capacity and runtime.
  • Grid resolution schedule = G=8 then 40 then 48
    Follows the KAN training protocol from [4]; selected by hand and not systematically swept in the main model.
  • Optimizer hyperparameters = lr=0.005, weight_decay=1e-5, batch_size=30000
    Set by hand; standard choices, but still free choices that affect results.
  • Reflectivity feature statistics = min, max, mean, var, dispersion, radial measures
    Handcrafted descriptors in Section III-F; the specific set is chosen by the authors and is not derived from a theory.
assumptions (5)
  • standard math Kolmogorov–Arnold superposition theorem
    Invoked in Section III-B to justify representing multivariate functions as compositions of univariate functions.
  • standard math B-spline basis functions are universal on compact intervals
    Used in Section III-D to justify that each ϕ_ij can approximate any sufficiently smooth univariate function.
  • domain assumption Traversability is well-modeled as binary classification of polar cells using handcrafted geometric/radiometric features
    Section III-A formulates the problem this way; assumes the handcrafted features are sufficient for the task.
  • domain assumption Semantic labels (drivable/flat vs obstacles/vegetation/vehicles) are a valid ground-truth proxy for traversability
    Section IV-A derives labels from semantic annotations; this ignores robot-specific capabilities.
  • domain assumption Reflectivity is a calibrated, geometry-independent surface property distinct from remission
    Section III-F asserts this distinction without empirical validation; the physical basis is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TravKAN: Fast and Interpretable Nonlinear Traversability Analysis with Kolmogorov-Arnold Networks." pith.science (2026). https://pith.science/paper/PVILWLK5

@misc{pith2026260802320,
  author       = {Pith},
  title        = {Pith review of: TravKAN: Fast and Interpretable Nonlinear Traversability Analysis with Kolmogorov-Arnold Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PVILWLK5}},
  note         = {Machine review of arXiv:2608.02320}
}
read the original abstract

Traversability analysis is a fundamental capability for autonomous mobile robots operating in unstructured environments. While modern machine learning approaches such as deep neural networks and gradient-boosted trees achieve strong predictive performance, they lack interpretability and provide limited insight into the underlying terrain-robot interaction dynamics. In this paper, we propose TravKAN, a Kolmogorov-Arnold Network-based framework for fast, scalable, and interpretable traversability estimation. TravKAN represents multivariate decision functions through compositions of learnable univariate functions, enabling compact architectures and symbolic extraction of analytic expressions after training. In addition, we introduce a novel set of handcrafted features derived from the reflectivity channel of LiDAR sensors. To the best of our knowledge, reflectivity has not been systematically exploited for handcrafted traversability descriptors, despite its potential to capture material and surface properties complementary to geometric cues. We evaluate TravKAN on public, real-world urban and off-road datasets and compare it against strong baselines. TravKAN achieves strong performance across all metrics, outperforming conventional deep models and approaching the performance of XGBoost. TravKAN-Lite, i.e., TravKAN's symbolic representation, reveals meaningful nonlinear feature interactions and provides a compact, deployment-friendly, and fast analytic model. Ablation studies further show the robustness of our method to architectural variations and quantify the contribution of the proposed reflectivity-based features. These properties make TravKAN attractive for robotic systems requiring transparency, real-time computational efficiency, and interpretability in safety-critical decision-making.

Figures

Figures reproduced from arXiv: 2608.02320 by the authors.

Figure 1
Figure 1. By learning nonlinear transformations directly on [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. By learning nonlinear transformations directly on edges and parameterizing them with B-splines, TravKAN captures [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 1 canonical work pages

  1. [1]

    Learning traversability from point clouds in challenging scenarios,

    M. Bellone, G. Reina, L. Caltagirone, and M. Wahde, “Learning traversability from point clouds in challenging scenarios,”IEEE Trans. on Intelligent Transportation Systems (TITS), vol. 19, no. 1, pp. 296–305, 2018

    2018

  2. [2]

    Attention Is All You Need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention Is All You Need,” Advances in Neural Information Processing Systems, 2017

    2017

  3. [3]

    Xgboost: A scalable tree boosting system,

    T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting system,” inProc. of the Intl. Conf. on Knowledge Discovery and Data Mining, 2016, pp. 785–794

    2016

  4. [4]

    Kan: Kolmogorov-arnold networks,

    Z. Liu, Y . Wang, S. Vaidya, F. Ruehle, J. Halverson, M. Solja ˇci´c, T. Y . Hou, and M. Tegmark, “Kan: Kolmogorov-arnold networks,” arXiv preprint arXiv:2404.19756, 2024

    arXiv 2024

  5. [5]

    SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences,

    J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, and J. Gall, “SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences,” inProc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2019. [Online]. Available: https://arxiv.org/pdf/1904.01416.pdf

    arXiv 2019

  6. [6]

    nuScenes: A Multimodal Dataset for Autonomous Driving,

    H. Caesar, V . Bankiti, A. Lang, S. V ora, V . Liong, Q. Xu, A. Krishnan, Y . Pan, G. Baldan, and O. Beijbom, “nuScenes: A Multimodal Dataset for Autonomous Driving,” inProc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020

    2020

  7. [7]

    Available: https://doi.org/10.5281/zenodo.17243765

    [Online]. Available: https://doi.org/10.5281/zenodo.17243765

    work page doi:10.5281/zenodo.17243765

  8. [8]

    Pyramidal 3d feature fusion on polar grids for fast and robust traversability analysis on cpu,

    D. Fusaro, E. Olivastri, I. Donadi, D. Evangelista, E. Menegatti, and A. Pretto, “Pyramidal 3d feature fusion on polar grids for fast and robust traversability analysis on cpu,”Journal on Robotics and Autonomous Systems (RAS), vol. 170, p. 104524, 2023

    2023

Show all 42 references
  1. [9]

    Terrain classification using tof sensors for the enhancement of agricultural machinery traversability,

    F. Y . Narv ´aez, E. Gregorio, A. Escol `a, J. R. Rosell-Polo, M. Torres- Torriti, and F. A. Cheein, “Terrain classification using tof sensors for the enhancement of agricultural machinery traversability,”Journal of Terramechanics, vol. 76, pp. 1–13, 2018

    2018

  2. [10]

    Self-supervised learning to visually detect terrain surfaces for autonomous robots operating in forested terrain,

    S. Zhou, J. Xi, M. W. McDaniel, T. Nishihata, P. Salesses, and K. Iagnemma, “Self-supervised learning to visually detect terrain surfaces for autonomous robots operating in forested terrain,”Journal of Field Robotics, vol. 29, no. 2, pp. 277–297, 2012

    2012

  3. [11]

    P-SVM2: Enhancing LiDAR-based Traversability Analysis with Aug- mented Point Cloud Descriptor for Autonomous Mobile Systems,

    W. Li, D. Fusaro, E. Olivastri, S. Mosco, N. Bellotto, and A. Pretto, “P-SVM2: Enhancing LiDAR-based Traversability Analysis with Aug- mented Point Cloud Descriptor for Autonomous Mobile Systems,” in Proc. of the IEEE Intl. Conf. on Cybern. and Intelligent Systems (CIS), Robot...

    2024

  4. [12]

    Self-supervised classification for planetary rover terrain sensing,

    C. A. Brooks and K. D. Iagnemma, “Self-supervised classification for planetary rover terrain sensing,” inIEEE Aerosp. Conf., 2007, pp. 1–9

    2007

  5. [13]

    Traversability analysis for mobile robots in outdoor environments: A semi-supervised learning approach based on 3d-lidar data,

    B. Suger, B. Steder, and W. Burgard, “Traversability analysis for mobile robots in outdoor environments: A semi-supervised learning approach based on 3d-lidar data,” inProc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2015, pp. 3941–3946

    2015

  6. [14]

    Traversability classification for ugv navigation: a comparison of patch and superpixel representations,

    D. Kim, S. M. Oh, and J. M. Rehg, “Traversability classification for ugv navigation: a comparison of patch and superpixel representations,” inProc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2007, pp. 3166–3173

    2007

  7. [15]

    Lsk3dnet: Towards effective and efficient 3d perception with large sparse kernels,

    T. Feng, W. Wang, F. Ma, and Y . Yang, “Lsk3dnet: Towards effective and efficient 3d perception with large sparse kernels,” inProc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 14 916–14 927

    2024

  8. [16]

    Point transformer v3: Simpler faster stronger,

    X. Wu, L. Jiang, P.-S. Wang, Z. Liu, X. Liu, Y . Qiao, W. Ouyang, T. He, and H. Zhao, “Point transformer v3: Simpler faster stronger,” inProc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 4840–4851

    2024

  9. [17]

    Exploiting local features and range images for small data real-time point cloud semantic segmentation,

    D. Fusaro, S. Mosco, E. Menegatti, and A. Pretto, “Exploiting local features and range images for small data real-time point cloud semantic segmentation,” inProc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2024, pp. 4980–4987

    2024

  10. [18]

    2dpass: 2d priors assisted semantic segmentation on lidar point clouds,

    X. Yan, J. Gao, C. Zheng, C. Zheng, R. Zhang, S. Cui, and Z. Li, “2dpass: 2d priors assisted semantic segmentation on lidar point clouds,” inProc. of the Europ. Conf. on Computer Vision (ECCV), 2022, pp. 677–695

    2022

  11. [19]

    Rangenet++: Fast and accurate lidar semantic segmentation,

    A. Milioto, I. Vizzo, J. Behley, and C. Stachniss, “Rangenet++: Fast and accurate lidar semantic segmentation,” inProc. of the IEEE/RSJ Intl. Conf. on Int. Robots and Systems (IROS),, 2019, pp. 4213–4220

    2019

  12. [20]

    Lite-hdseg: Lidar semantic segmentation using lite harmonic dense convolutions,

    R. Razani, R. Cheng, E. Taghavi, and L. Bingbing, “Lite-hdseg: Lidar semantic segmentation using lite harmonic dense convolutions,” in Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2021, pp. 9550–9556

    2021

  13. [21]

    Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud,

    B. Wu, X. Zhou, S. Zhao, X. Yue, and K. Keutzer, “Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud,” inProc. of the IEEE Intl. Conf. on Robotics & Autom. (ICRA), 2019, pp. 4376–4382

    2019

  14. [22]

    Polarnet: An improved grid representation for online lidar point clouds semantic segmentation,

    Y . Zhang, Z. Zhou, P. David, X. Yue, Z. Xi, B. Gong, and H. Foroosh, “Polarnet: An improved grid representation for online lidar point clouds semantic segmentation,” inProc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 9598– 9607

    2020

  15. [23]

    4d spatio-temporal convnets: Minkowski convolutional neural networks,

    C. Choy, J. Gwak, and S. Savarese, “4d spatio-temporal convnets: Minkowski convolutional neural networks,” inProc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 3070–3079

    2019

  16. [24]

    3d semantic segmen- tation with submanifold sparse convolutional networks,

    B. Graham, M. Engelcke, and L. v. d. Maaten, “3d semantic segmen- tation with submanifold sparse convolutional networks,” inProc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 9224–9232

    2018

  17. [25]

    Foresttrav: 3d lidar-only forest traversability estimation for au- tonomous ground vehicles,

    F. A. Ruetz, N. Lawrance, E. Hern ´andez, P. V . Borges, and T. Peynot, “Foresttrav: 3d lidar-only forest traversability estimation for au- tonomous ground vehicles,”IEEE Access, vol. 12, pp. 37 192–37 206, 2024

    2024

  18. [26]

    Pointnet: Deep learning on point sets for 3d classification and segmentation,

    R. Q. Charles, H. Su, M. Kaichun, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” inProc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 77–85

    2017

  19. [27]

    Pointnet++: Deep hierar- chical feature learning on point sets in a metric space,

    C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “Pointnet++: Deep hierar- chical feature learning on point sets in a metric space,” inAdvances in Neural Information Processing Systems, vol. 30, 2017

    2017

  20. [28]

    Randla-net: Efficient semantic segmentation of large- scale point clouds,

    Q. Hu, B. Yang, L. Xie, S. Rosa, Y . Guo, Z. Wang, N. Trigoni, and A. Markham, “Randla-net: Efficient semantic segmentation of large- scale point clouds,” inProc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 11 105–11 114

    2020

  21. [29]

    Kpconv: Flexible and deformable convolution for point clouds,

    H. Thomas, C. R. Qi, J.-E. Deschaud, B. Marcotegui, F. Goulette, and L. Guibas, “Kpconv: Flexible and deformable convolution for point clouds,” inProc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2019, pp. 6410–6419

    2019

  22. [30]

    Gonet: A semi-supervised deep learning approach for traversability estimation,

    N. Hirose, A. Sadeghian, M. V ´azquez, P. Goebel, and S. Savarese, “Gonet: A semi-supervised deep learning approach for traversability estimation,” inProc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2018, pp. 3044–3051

    2018

  23. [31]

    Vunet: Dynamic scene view synthesis for traversability estimation using an rgb camera,

    N. Hirose, A. Sadeghian, F. Xia, R. Mart ´ın-Mart´ın, and S. Savarese, “Vunet: Dynamic scene view synthesis for traversability estimation using an rgb camera,”IEEE Robotics and Automation Letters (RA-L), vol. 4, no. 2, pp. 2062–2069, 2019

  24. [32]

    Deep visual mpc-policy learning for navigation,

    N. Hirose, F. Xia, R. Mart ´ın-Mart´ın, A. Sadeghian, and S. Savarese, “Deep visual mpc-policy learning for navigation,”IEEE Robotics and Automation Letters (RA-L), vol. 4, no. 4, pp. 3184–3191, 2019

    2019

  25. [33]

    Adapnet: Adap- tive semantic segmentation in adverse environmental conditions,

    A. Valada, J. Vertens, A. Dhall, and W. Burgard, “Adapnet: Adap- tive semantic segmentation in adverse environmental conditions,” in Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2017, pp. 4644–4651

    2017

  26. [34]

    Deep reinforcement learning for safe local planning of a ground vehicle in unknown rough terrain,

    S. Josef and A. Degani, “Deep reinforcement learning for safe local planning of a ground vehicle in unknown rough terrain,”IEEE Rob. and Autom. Letters (RA-L), vol. 5, no. 4, pp. 6748–6755, 2020

    2020

  27. [35]

    Learning to drive off road on smooth terrain in unstructured environments using an on-board camera and sparse aerial images,

    T. Manderson, S. Wapnick, D. Meger, and G. Dudek, “Learning to drive off road on smooth terrain in unstructured environments using an on-board camera and sparse aerial images,” inProc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2020, pp. 1263–1269

    2020

  28. [36]

    Learning-based end-to-end path planning for lunar rovers with safety constraints,

    X. Yu, P. Wang, and Z. Zhang, “Learning-based end-to-end path planning for lunar rovers with safety constraints,”Sensors, vol. 21, no. 3, 2021

    2021

  29. [37]

    Robot navigation of environments with unknown rough terrain using deep reinforcement learning,

    K. Zhang, F. Niroui, M. Ficocelli, and G. Nejat, “Robot navigation of environments with unknown rough terrain using deep reinforcement learning,” inProc. of the IEEE Intl. Symp. on Safety, Security, and Rescue Robotics (SSRR), 2018, pp. 1–7

    2018

  30. [38]

    BADGR: An Autonomous Self- Supervised Learning-Based Navigation System,

    G. Kahn, P. Abbeel, and S. Levine, “BADGR: An Autonomous Self- Supervised Learning-Based Navigation System,”IEEE Robotics and Automation Letters (RA-L), vol. 6, no. 2, pp. 1312–1319, 2021

    2021

  31. [39]

    Incorporating multi-context into the traversability map for urban autonomous driving using deep inverse reinforcement learning,

    C. Jung and D. H. Shim, “Incorporating multi-context into the traversability map for urban autonomous driving using deep inverse reinforcement learning,”IEEE Robotics and Automation Letters (RA- L), vol. 6, no. 2, pp. 1662–1669, 2021

    2021

  32. [40]

    Off-road autonomous vehicles traversability analysis and trajectory planning based on deep inverse reinforcement learning,

    Z. Zhu, N. Li, R. Sun, D. Xu, and H. Zhao, “Off-road autonomous vehicles traversability analysis and trajectory planning based on deep inverse reinforcement learning,” inProc. of the IEEE Vehicles Sympo- sium (IV), 2020, pp. 971–977

    2020

  33. [41]

    An autonomous exploration algorithm using environment-robot interacted traversability analysis,

    Y . Tang, J. Cai, M. Chen, X. Yan, and Y . Xie, “An autonomous exploration algorithm using environment-robot interacted traversability analysis,” inProc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2019, pp. 4885–4890

    2019

  34. [42]

    Reflectivity is all you need!: Advancing lidar semantic segmentation,

    K. Viswanath, P. Jiang, and S. Saripalli, “Reflectivity is all you need!: Advancing lidar semantic segmentation,”IFAC-PapersOnLine, vol. 59, no. 3, pp. 43–48, 2025

    2025

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.