Pith. sign in

REVIEW 3 major objections 8 minor 39 references

Overhead maps plus human GPS traces can teach a robot long-horizon off-road routes that stay within a few percent of human path length.

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 ↗

Overhead multi-modal learning with PU human-trajectory supervision and LiDAR priors yields global off-road costmaps that nearly match human path length and sharply cut interventions versus local planners.

T0 review reviewed 2026-07-30 challenge →

load-bearing objection Solid systems paper with a real dataset and closed-loop Warthog runs, but the abstract’s 3.66%/85% numbers do not match Table III and the local-only baseline is missing on the long routes. the 3 major comments →

arxiv 2607.23743 v1 pith:7Q6QQ45E submitted 2026-07-26 cs.RO

Learning Traversability-Aware Global Planners for Long Horizon Off-Road Navigation

Kasi Viswanath , Jason M. Gregory , Shaunak Kolhe , Srikanth Saripalli This is my paper
classification cs.RO
keywords off-road navigationtraversability estimationglobal planningpositive-unlabeled learningaerial LiDARsatellite imageryhuman demonstrationscostmaps
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.

The reading

Off-road robots only see a few meters ahead, but good routes depend on slopes, water, trails, and vegetation far beyond that horizon. This paper shows that satellite imagery, aerial LiDAR, and map vectors can be turned into a continuous traversability map when the only dense supervision is human-driven GPS tracks, treated as positive corridors rather than full labels, and shaped by simple geometric priors from LiDAR. The same model also predicts goal-conditioned path likelihood, so a global planner can propose routes that match how people actually drive. The authors release a large multi-modal dataset and demonstrate on a wheeled robot that the resulting long-range plans cut operator interventions by roughly 85% versus local planning alone while staying close to human path length.

Core claim

A dual-stream network can learn a dense continuous traversability map and goal-conditioned path likelihood from overhead satellite, LiDAR, and OSM inputs when supervised by sparse human GPS trajectories under a positive-unlabeled corridor loss, plus weak OSM/trajectory anchors and self-supervised LiDAR priors on intensity, slope, and height gradient—without dense labels or a differentiable planner—and that map supports long-horizon plans that track human routes and sharply reduce interventions in real field trials.

What carries the argument

Positive-unlabeled path-corridor loss on human GPS bands (with mass and total-variation regularizers), fused multi-modal features via spatially varying FiLM, and LiDAR geometric priors that shape the continuous traversability head so unvisited cells stay unlabeled while hazardous geometry is penalized.

Load-bearing premise

Human driving bands plus hand-shaped LiDAR priors are enough to rank real wheeled feasibility everywhere, even though most terrain was never driven and never labeled as bad.

What would settle it

On held-out start–goal pairs in terrain unlike the training demos, if A* on the learned costmap systematically routes through slopes, water, or vegetation a human would refuse—or forces large path-length overhead and frequent interventions versus local planning—the central claim fails.

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

If this is right

  • Global off-road planners can be trained from public overhead data and GPS traces without hand-drawn costmaps or dense semantic labels.
  • Long-range waypoint sequences from the learned map measurably outperform local-planner-only autonomy on path efficiency and intervention rate.
  • The same costmap can yield multiple competitive K-shortest routes with meaningful cost-per-distance differences for operator choice.
  • When onboard sensing finds an unmapped obstacle, injecting a local high-cost patch and replanning on the global map is enough to recover without abandoning the prior.

Where Pith is reading between the lines

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

  • Seasonal vegetation and surface change will drift the static overhead prior, so online fusion with live perception is likely required before high-speed deployment.
  • The method’s conservatism in ambiguous brush may be a direct cost of missing negative labels; adding sparse failure or intervention labels could shrink the human–robot length gap.
  • The released multi-scene coverage makes cross-region transfer and embodiment-specific prior retuning testable at scale.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 8 minor

Summary. The paper presents a dual-stream, multi-modal network that learns a continuous overhead traversability map and a goal-conditioned path-likelihood map from satellite imagery (frozen DINOv3-SAT encoder), rasterized aerial LiDAR (height/slope/intensity), and OSM priors, fused via spatial FiLM. Supervision combines a Positive-Unlabeled (nnPU) path-corridor loss on human GPS trajectories with weak positive anchors (OSM tracks, aggregated trajectories) and self-supervised LiDAR geometric priors (intensity, Gaussian slope, height-gradient), avoiding dense manual labels and differentiable planners. The authors release a dataset of 299 scenes (~1,244 km², 1,130 km of GPS traces). Evaluation comprises (i) a held-out-scene costmap benchmark (n=356 trajectories, 20 test scenes) where A* on the learned costmap reproduces human paths better than height-map, Trailblazer, and OVerSeeC baselines (Fréchet 61.3 m vs 77.5–108.4 m); (ii) loss ablations showing increasing corridor-vs-alternative discrimination; and (iii) field trials on a Clearpath Warthog over seven routes at two sites, claiming 3.66% path-length overhead vs human teleoperation and ~85–90% fewer interventions than local-planner-only (SRP) autonomy.

Significance. If the results hold, the work makes two contributions of clear value to the field-robotics community. First, the released dataset — 299 co-registered scenes spanning ~1,244 km² with satellite, LiDAR, OSM, and 1,130 km of human driving — is, to my knowledge, the largest geographic-coverage resource of its kind and is publicly released with a concrete link; this alone is likely to be used by others. Second, the PU formulation for path supervision plus self-supervised LiDAR priors is a sensible, well-motivated answer to a real problem (no dense traversability labels at scale), and the ablation in Table II gives a falsifiable, quantitative account of what each loss group contributes. The held-out-scene benchmark with an external planner (A*) as the evaluation harness is a sound protocol design, and the real-robot deployment on a Warthog with a full autonomy stack (DLIO, MPPI/TerrainNet) demonstrates system-level integration rather than a tabletop result. The qualitative direction of the field results — global planner routes more safely than local-only autonomy — is plausible and consistent with the qualitative evidence in Figs. 5–6.

major comments (3)
  1. [§VII, Table III (also Abstract and §I contributions)] The headline field-trial numbers are not derivable from the paper's own data. Per-route LRP distance overheads computed from Table III are 4.7%, 1.8%, 5.9%, 17.0%, 8.1%, 2.0%, 16.4%: a route-mean of 8.0% (7.4% on the four routes with SRP data) and a distance-pooled ratio of 5964/5506 = 8.3%. No standard aggregation of Table III yields the stated 3.66%. Similarly, SRP overheads average 11.6% (pooled 12.9%), not the stated 9.7%. The intervention statistics are also inconsistent: Table III gives 8 LRP interventions over 5964 m, i.e., one per ~745 m, yet §VII states one per 480 m; the SRP figure (14 over 1979 m ≈ one per 141 m) does match the three-stated 140 m. The implied intervention reduction is ~71% from the claimed rates and ~81% from the table-derived rates, while the paper variously claims 85% (abstract, contributions) and 90% (§VII). These numbers carry the abstract; the authors mus
  2. [§VII, Table III and Q2 discussion] The comparative claim 'global planner outperforms local-planner-only autonomy' is supported only on Routes 1–4 (≤765 m); SRP-only baselines are absent on Routes 5–7, which include the two 1680 m missions that instantiate the paper's stated long-horizon motivation. The safety rationale for omitting SRP on hazardous/long routes is understandable and honestly disclosed, but the consequence is that the central comparison carries no baseline exactly where it should bite. Additionally, each route appears to be a single run: no repetitions, no variance, and the intervention criterion ('obstacle encounters or large deviations') is operator-judgment-dependent. At minimum the paper needs (a) explicit acknowledgment that the long-horizon routes are uncontrolled demonstrations, not comparisons; (b) repeated runs or some variance estimate on at least a subset of routes; and ideally (c) a surrogate lo
  3. [§V.B, Table II] The ablation's key metric — mean traversability along 'dense human corridors' vs 'sparse alternative routes' between the same start–goal pairs — is under-specified. How are sparse alternatives generated (manual annotation, synthetic path sampling, low-density GPS cells)? How many routes, from how many scenes, and is there overlap with training scenes? Because the corridor scores approach saturation (0.980 in Variant D), the ∆ metric is partly a ceiling effect; the paper should report the number of route pairs, per-scene breakdown, and ideally a significance test. As written, Table II is the main quantitative evidence that the LiDAR priors (the paper's mechanism for penalizing hazardous geometry without negative labels) actually work, so its protocol needs to be reproducible.
minor comments (8)
  1. [§V.A, Table I] The benchmark metric rewards alignment with human trajectories, which is exactly the quantity the proposed method is trained to match (via PU supervision on human GPS bands), whereas the height-map and OVerSeeC baselines are not trained for human alignment. Held-out scenes mitigate but do not remove this asymmetry; a sentence acknowledging it, and ideally one metric not defined by human imitation (e.g., rate of A* paths intersecting hand-labeled hazards), would strengthen the section.
  2. [§VII, Table III, Route 4] On Route 4 the SRP-only run (326 m) is shorter than the LRP run (338 m) and has only one intervention; this route cuts against the narrative and deserves explicit comment, especially since Route 4 is also LRP's worst overhead (17%).
  3. [§IV.C, Eq. (4)] The nnPU loss is stated with β=0 and γ=1; with β=0 the non-negative correction branch (Δ < −β) reduces to Δ < 0, i.e., standard uPU-style clamping at zero. Please clarify whether this is intended, and how the prior π is estimated from the training set (corridor pixel fraction?), since π enters the risk estimator linearly.
  4. [§IV.C, Eqs. (6)–(11)] Several hand-set constants lack sensitivity analysis: T* = 0.75 (Eq. 7), σ_s = 0.445, σ_h, g_95 normalization, and the eight loss weights 'chosen empirically.' A small sensitivity sweep for at least σ_s and T* would substantially increase confidence that the priors are not mis-ranking wheeled feasibility.
  5. [§VII, Table IV] The 'cost / distance (×10⁻¹)' column has no stated units or normalization (cost is in [0,1] per cell via C = 1−T, but path cost depends on grid resolution G_res, which varies by scene). Also the sentence 'demonstrating that the preservation operator choice' is garbled.
  6. [§VII] Grammar: 'Where did the LRP fell short of the human?' → 'fall short'. The abstract and contributions say 85% intervention reduction while §VII says 90%; please make consistent (see major comment 1).
  7. [Fig. 3 and Fig. 5] Costmap figures lack colorbars/units, making the claimed 'progressive improvement' in Fig. 3 hard to assess quantitatively. Figure 6 would benefit from start/goal markers and per-route labels directly on the panels.
  8. [§V] Training details are thin: optimizer, learning rate, epochs, batch size, and data augmentation are not reported, and the dataset split (7533/1716 samples) is given in samples rather than scenes — please state the number of distinct scenes in train vs validation to substantiate the 'completely separate scenes' claim.

Circularity Check

1 steps flagged

No load-bearing circularity: human-demo PU supervision makes offline human-path alignment expected, but traversability priors, held-out scenes, and SRP field baselines are independent of that objective.

specific steps
  1. other [§IV.C Path Corridor Loss (Eqs. 2–5); §V.A Benchmark protocol and Table I]
    "Path supervision follows a Positive–unlabeled (PU) formulation [7]: pixels within a band of radius r=5px around human-driven GPS trajectories are treated as positives P... We therefore evaluate costmaps by how closely an A* planner, run on each costmap, reproduces human-driven reference trajectories: a costmap is better if it leads the planner to paths a human operator would have chosen."

    The path head is optimized to raise likelihood on human GPS corridors; the primary offline benchmark then ranks costmaps by how well A* on them recovers those same human trajectories (d_F, r_c). That coupling is expected under imitation learning and is mitigated by held-out scenes, separate LiDAR/OSM shaping of T̂, and physical SRP baselines—so it is mild evaluation tautology, not a definitional or fitted-parameter reduction of the field result.

full rationale

This is an engineering/learning paper, not a first-principles derivation. The path-logits head is explicitly trained with a Positive-Unlabeled corridor loss on human GPS bands (Eqs. 2–5), so offline metrics that score A* paths by Fréchet distance and cost ratio to human tracks (§V.A, Table I) partly measure success at the training task—standard imitation-learning evaluation on held-out scenes, not Eq-X=Eq-Y by construction. Traversability is shaped by separate self-supervised LiDAR priors (intensity, Gaussian slope, height-gradient; Eqs. 8–10) and weak OSM/trajectory anchors, ablated in Table II, and converted to cost via C=1−T for external A*/Yen planning. Field claims compare LRP+SRP against SRP-only and human teleop on a physical Warthog (§VII, Table III), which is not forced by the training loss. Self-citation of Trailblazer [27] is only as a costmap baseline, not a uniqueness theorem or ansatz smuggled into the derivation. No fitted scalar is renamed a prediction; loss weights and σ_s are acknowledged empirical choices. Score 1 reflects only the mild, expected coupling of demo supervision to human-alignment offline metrics—not circular reduction of the central system claim.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 2 invented entities

The claim rests on standard planning/learning tools plus domain choices that human GPS corridors and simple LiDAR Gaussians adequately supervise dense traversability, and that static overhead costmaps plus local MPPI suffice for the reported intervention gains. Many scalar loss and prior knobs are set by hand; no new physical entity is postulated beyond the learned map heads.

free parameters (8)
  • Path and auxiliary loss weights (λ_p, λ_m, λ_tv, λ_trk, λ_tj, λ_int, λ_sl, λ_ht) = λ_p=1.0; auxiliaries/priors λ≤0.2
    Stated as chosen empirically and fixed; path weight 1.0 with auxiliaries ≤0.2 dominate the trained objective (§IV.C.4).
  • Slope prior σ_s = 0.445
    Sets Gaussian traversability crossover near 30° for wheeled platforms; hand-chosen geometry prior (§IV.C.3).
  • Height-gradient prior σ_h and g_95 normalization = σ_h unspecified; g_95 = 95th percentile
    Controls penalty on elevation discontinuities; σ_h not numerically fixed in text; g_95 is data percentile (§IV.C.3).
  • Trajectory anchor T* and OSM track target = T*=0.75; OSM target=1
    Soft positive traversability targets on historical paths (0.75) and OSM geometries (1.0) (§IV.C.2).
  • PU corridor radius r, π prior, nnPU β/γ = r=5 px; β=0; γ=1; π offline
    Defines positive band around GPS and non-negative PU correction; π estimated offline from training set (§IV.C.1).
  • Auxiliary loss warmup α_aux over E_warm = E_warm=3 epochs
    Linear ramp so path loss dominates early training (§IV.C.4).
  • Grid resolution rule and fusion size = ≥15 pts/cell; 256 fusion / 1024 out
    G_res from ≥15 pts/cell; fusion at 256×256, outputs 1024×1024—design choices affecting labels and metrics (§IV.A–B).
  • A* multi-candidate K and cost/length weights α,β; C=1−T = K=4 in benchmark; α,β weighted objective
    Downstream planner knobs used in benchmark and deployment (§V, §VI).
axioms (6)
  • domain assumption Unvisited pixels must not be treated as negatives; PU learning with human corridor positives is a valid supervision model for path likelihood.
    Core of L_path (§IV.C.1); without it the labeling story collapses.
  • domain assumption LiDAR intensity complement, slope, and height-gradient Gaussians are faithful enough self-supervised proxies for wheeled traversability hazards.
    Lint, Lslope, Lheight supply the only dense negative-leaning terrain signal (§IV.C.3).
  • domain assumption Frozen DINOv3-SAT features plus raster LiDAR/OSM suffice as multi-modal inputs for land-cover and geometry relevant to off-road planning.
    Architecture premise (§IV.B.1–4).
  • domain assumption Costmap quality is well measured by how closely A* paths match human references (Fréchet, cost ratio, length ratio).
    Explicit evaluation protocol (§V.A); embodiment- and preference-dependent.
  • domain assumption Static overhead maps plus local TerrainNet/MPPI execution represent the deployed autonomy comparison fairly when SRP-only is safety-limited on some routes.
    Field protocol §VI–VII; affects intervention claims.
  • standard math Standard nnPU risk estimator, A*, Yen K-shortest paths, and FiLM fusion behave as in the cited literature.
    Used as building blocks without re-proof (§IV–VI).
invented entities (2)
  • Goal-conditioned dual-head overhead traversability and path-likelihood model (LRP costmap) independent evidence
    purpose: Produce dense T∈[0,1] and path logits from satellite+LiDAR+OSM+start/goal for global waypoint planning.
    Learned representation rather than a new physical object; existence is operational (network outputs), validated by ablations and field use.
  • Off-road global navigation dataset (299 scenes, ~1244 km², 1130 km GPS) independent evidence
    purpose: Provide paired multi-modal overhead inputs and human traces for training/evaluation at scale.
    Curated resource claimed released; independent use depends on actual public availability of the HF artifacts.

reviewed 2026-07-30 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Traversability-Aware Global Planners for Long Horizon Off-Road Navigation." pith.science (2026). https://pith.science/paper/7Q6QQ45E

@misc{pith2026260723743,
  author       = {Pith},
  title        = {Pith review of: Learning Traversability-Aware Global Planners for Long Horizon Off-Road Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7Q6QQ45E}},
  note         = {Machine review of arXiv:2607.23743}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Autonomous navigation across large off-road environments remains a challenging problem. Onboard sensors perceive only the immediate surroundings, yet safe and efficient routes depend on terrain features that extend well beyond the sensor horizon. Geo-spatial data sources such as satellite imagery, aerial LiDAR, and vector maps can close this gap, but learning traversability from them is difficult: dense labels are unavailable at scale, and existing methods rely on short-range sensing. We propose an efficient formulation that learns a continuous traversability map from overhead data, supervised directly by human-driven GPS trajectories and shaped by self-supervised geometric priors from LiDAR. Alongside the model, we release a public dataset of 299 scenes spanning $\sim\!1{,}244\,\mathrm{km}^{2}$ of diverse terrain, paired with $1{,}130\,\mathrm{km}$ of human driving. In field trials on a Clearpath Warthog across seven routes at two sites, our method achieves trajectories within $3.66\%$ of human path length and reduces operator interventions by $\sim\!85\%$ compared to local-planner-only autonomy.

Figures

Figures reproduced from arXiv: 2607.23743 by Jason M. Gregory, Kasi Viswanath, Shaunak Kolhe, Srikanth Saripalli.

Figure 1
Figure 1. Figure 1: Warthog in action across diverse off-road terrain, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the architecture. Pink block denote frozen pretrained modules (DINOv3-SAT ViT-L encoder); blue [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ablation traversability maps on a held-out desert [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sensor suite integrated on the Warthog platform [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overview of the Levee test site (a, b), showing [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Full trajectory comparison across all seven routes, [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

39 extracted references · 1 canonical work pages

  1. [1]

    A survey of motion planning and control techniques for self-driving urban vehicles,

    B. Paden, M. C ´ap, S. Z. Yong, D. Yershov, and E. Frazzoli, “A survey of motion planning and control techniques for self-driving urban vehicles,”IEEE Transactions on Intelligent Vehicles, vol. 1, no. 1, pp. 33–55, 2016. [Online]. Available: https://arxiv.org/abs/1604.07446

    Pith/arXiv arXiv 2016

  2. [2]

    Winning the darpa grand challenge with an ai robot,

    M. Montemerlo, S. Thrun, H. Dahlkamp, D. Stavens, and S. Stro- hband, “Winning the darpa grand challenge with an ai robot,” inAAAI, 2006

    2006

  3. [3]

    Obstacle de- tection and terrain classification for autonomous off-road navigation,

    R. Manduchi, A. Castano, A. Talukder, and L. Matthies, “Obstacle de- tection and terrain classification for autonomous off-road navigation,” Autonomous Robots, vol. 18, no. 1, pp. 81–102, 2005

    2005

  4. [4]

    Towards learned traversability for robot navigation: From underfoot to the far field,

    A. Howard, M. Turmon, L. Matthies, B. Tang, A. Angelova, and E. Mjolsness, “Towards learned traversability for robot navigation: From underfoot to the far field,”Journal of Field Robotics, vol. 23, no. 11-12, pp. 1005–1017, 2006. [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1002/rob.20168

    work page doi:10.1002/rob.20168 2006

  5. [5]

    Large-scale cost function learning for path planning using deep inverse reinforcement learning,

    M. Wulfmeier, D. Rao, D. Z. Wang, P. Ondruska, and I. Posner, “Large-scale cost function learning for path planning using deep inverse reinforcement learning,”The International Journal of Robotics Research, vol. 36, no. 10, pp. 1073–1087, 2017

    2017

  6. [6]

    Terp: Reliable planning in uneven outdoor environments using deep reinforcement learning,

    K. Weerakoon, A. J. Sathyamoorthy, U. Patel, and D. Manocha, “Terp: Reliable planning in uneven outdoor environments using deep reinforcement learning,” in2022 International Conference on Robotics and Automation (ICRA), 2022, pp. 9447–9453

    2022

  7. [7]

    Positive-unlabeled learning with non-negative risk estimator,

    R. Kiryo, G. Niu, M. du Plessis, and M. Sugiyama, “Positive-unlabeled learning with non-negative risk estimator,” inAdvances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  8. [8]

    A rugd dataset for autonomous navigation and visual perception in unstructured outdoor environments,

    M. Wigness, S. Eum, J. G. Rogers, D. Han, and H. Kwon, “A rugd dataset for autonomous navigation and visual perception in unstructured outdoor environments,” inInternational Conference on Intelligent Robots and Systems (IROS), 2019

    2019

  9. [9]

    Rellis-3d dataset: Data, benchmarks and analysis,

    P. Jiang, P. Osteen, M. Wigness, and S. Saripalli, “Rellis-3d dataset: Data, benchmarks and analysis,” in2021 IEEE International Confer- ence on Robotics and Automation (ICRA), 2021, pp. 1110–1116

    2021

  10. [10]

    The goose dataset for perception in unstructured environments,

    P. Mortimer, R. Hagmanns, M. Granero, T. Luettel, J. Petereit, and H.-J. Wuensche, “The goose dataset for perception in unstructured environments,” inProceedings of IEEE International Conference on Robotics and Automation (ICRA), 2024. [Online]. Available: https://arxiv.org/abs/2310.16788

    Pith/arXiv arXiv 2024

  11. [11]

    Deep multispectral semantic scene understanding of forested environments using multi- modal fusion,

    A. Valada, G. Oliveira, T. Brox, and W. Burgard, “Deep multispectral semantic scene understanding of forested environments using multi- modal fusion,” inInternational Symposium on Experimental Robotics (ISER), 2016

    2016

  12. [12]

    Go: The great outdoors multimodal dataset,

    P. Jiang, K. Viswanath, A. Nagariya, G. Chustz, M. Wigness, P. Osteen, T. Overbye, C. Ellis, L. Quang, and S. Saripalli, “Go: The great outdoors multimodal dataset,” 2025. [Online]. Available: https://arxiv.org/abs/2501.19274

    Pith/arXiv arXiv 2025

  13. [13]

    DigiForests: A Longitudinal LiDAR Dataset for Forestry Robotics,

    M. V . Malladi, N. Chebrolu, I. Scacchetti, L. Lobefaro, T. Guadagnino, B. Casseau, H. Oh, L. Freißmuth, M. Karppinen, J. Schweier, S. Leutenegger, J. Behley, C. Stachniss, and M. Fallon, “DigiForests: A Longitudinal LiDAR Dataset for Forestry Robotics,” 2025

    2025

  14. [14]

    Tartandrive 2.0: More modalities and better infrastructure to further self-supervised learning research in off-road driving tasks,

    M. Sivaprakasam, P. Maheshwari, M. G. Castro, S. Triest, M. Nye, S. Willits, A. Saba, W. Wang, and S. Scherer, “Tartandrive 2.0: More modalities and better infrastructure to further self-supervised learning research in off-road driving tasks,” 2024

    2024

  15. [15]

    World imagery [basemap],

    Esri, DigitalGlobe, GeoEye, Earthstar Geographics, CNES/Airbus DS, USDA, USGS, AeroGRID, IGN, and the GIS User Community, “World imagery [basemap],” 2026, accessed:

    2026

  16. [16]

    Copernicus sentinel-2 calibration and validation,

    E. S. A. (ESA), “Copernicus sentinel-2 calibration and validation,” European Journal of Remote Sensing, vol. 52, no. 1, pp. 253– 255, 2019. [Online]. Available: https://publications.jrc.ec.europa.eu/ repository/handle/JRC115729

    2019

  17. [17]

    3d elevation program (3DEP) - Lidar Point Cloud,

    U.S. Geological Survey, “3d elevation program (3DEP) - Lidar Point Cloud,” U.S. Geological Survey, 2024, accessed: 2026. [Online]. Available: https://apps.nationalmap.gov/lidar-explorer/

    2024

  18. [18]

    Open- topography: A services oriented architecture for community access to lidar topography,

    C. J. Crosby, J. R. Arrowsmith, V . Nandigam, and C. Baru, “Open- topography: A services oriented architecture for community access to lidar topography,” inProceedings of the 2nd International Conference on Computing for Geospatial Research and Applications. ACM, 2010

    2010

  19. [19]

    Planet dump retrieved from https://planet.osm.org ,

    OpenStreetMap contributors, “Planet dump retrieved from https://planet.osm.org ,” 2025, accessed: [2026]

    2025

  20. [20]

    Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification,

    P. Helber, B. Bischke, A. Dengel, and D. Borth, “Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 12, no. 7, pp. 2217–2226, 2019

    2019

  21. [21]

    Landsat 8-9 Operational Land Imager / Thermal Infrared Sensor level-2, collection 2,

    Earth Resources Observation and Science (EROS) Center, “Landsat 8-9 Operational Land Imager / Thermal Infrared Sensor level-2, collection 2,” 2020. [Online]. Available: https://doi.org/10.5066/ P9OGBGM6

    2020

  22. [22]

    Flair: a country-scale land cover semantic segmentation dataset from multi-source optical imagery,

    A. Garioud, N. Gonthier, L. Landrieu, A. D. Wit, M. Valette, M. Poup ´ee, S. Giordano, and B. Wattrelos, “Flair: a country-scale land cover semantic segmentation dataset from multi-source optical imagery,” inAdvances in Neural Information Processing Systems (NeurIPS) 2023, 2023

    2023

  23. [23]

    How does it feel? self-supervised costmap learning for off-road vehicle traversability,

    M. Guaman Castro, S. Triest, W. Wang, J. M. Gregory, F. Sanchez, J. G. Rogers III, and S. Scherer, “How does it feel? self-supervised costmap learning for off-road vehicle traversability,” IEEE, 2023

    2023

  24. [24]

    Salon: Self-supervised adaptive learning for off-road navigation,

    M. Sivaprakasam, S. Triest, C. Ho, S. Aich, J. Lew, I. Adu, W. Wang, and S. Scherer, “Salon: Self-supervised adaptive learning for off-road navigation,” 2024. [Online]. Available: https://arxiv.org/ abs/2412.07826

    Pith/arXiv arXiv 2024

  25. [25]

    High performance outdoor navigation from overhead data using imitation learning,

    D. Silver, J. A. Bagnell, and A. Stentz, “High performance outdoor navigation from overhead data using imitation learning,” inRobotics: Science and Systems IV, O. Brock, J. Trinkle, and F. Ramos, Eds. MIT Press, 2008. [Online]. Available: https://www.roboticsproceedings.org/rss04/p34.pdf

    2008

  26. [26]

    Overseec: Open-vocabulary costmap generation from satellite images and natural language,

    R. Rana, J. Quattrociocchi, D. Lee, C. Ellis, A. Adkins, A. Uccello, G. Warnell, and J. Biswas, “Overseec: Open-vocabulary costmap generation from satellite images and natural language,” 2026. [Online]. Available: https://arxiv.org/abs/2602.18606

    arXiv 2026

  27. [27]

    Trailblazer: Learning offroad costmaps for long range planning,

    K. Viswanath, F. Sanchez, T. Overbye, J. M. Gregory, and S. Saripalli, “Trailblazer: Learning offroad costmaps for long range planning,”

  28. [28]

    Path planning using neural a* search,

    R. Yonetani, T. Taniai, M. Barekatain, M. Nishimura, and A. Kanezaki, “Path planning using neural a* search,” inProceedings of the 38th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, M. Meila and T. Zhang, Eds., vol. 139. PMLR, 18–24 Jul 2021, pp. 12 029–12 039. [Online]. Available: http://proceedings.mlr.pres...

    2021

  29. [29]

    Sim ´eoni, H

    O. Sim ´eoni, H. V . V o, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V . Khalidov, M. Szafraniec, S. Yi, M. Ramamonjisoa, F. Massa, D. Haziza, L. Wehrstedt, J. Wang, T. Darcet, T. Moutakanni, L. Sentana, C. Roberts, A. Vedaldi, J. Tolan, J. Brandt, C. Couprie, J. Mairal, H. J ´egou, P. Labatut, and P. Bojanowski, “DINOv3,” 2025. [Online]. Available: h...

    2025

  30. [30]

    Film: visual reasoning with a general conditioning layer,

    E. Perez, F. Strub, H. de Vries, V . Dumoulin, and A. Courville, “Film: visual reasoning with a general conditioning layer,” inProceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, ser. AAAI’18/IAAI’18/EAAI’18. AAAI Press, 2018

    2018

  31. [31]

    A formal basis for the heuristic determination of minimum cost paths,

    P. Hart, N. Nilsson, and B. Raphael, “A formal basis for the heuristic determination of minimum cost paths,”IEEE Transactions on Systems Science and Cybernetics, vol. 4, no. 2, pp. 100–107, 1968. [Online]. Available: https://doi.org/10.1109/tssc.1968.300136

    arXiv 1968

  32. [32]

    Computing discrete fr ´echet distance,

    T. Eiter and H. Mannila, “Computing discrete fr ´echet distance,” in Proceedings of the Second International Conference on Computer Vision, Pattern Recognition and Image Processing. IEEE, 1994, pp. 313–316

    1994

  33. [33]

    Direct lidar-inertial odom- etry: Lightweight lio with continuous-time motion correction,

    K. Chen, R. Nemiroff, and B. T. Lopez, “Direct lidar-inertial odom- etry: Lightweight lio with continuous-time motion correction,”2023 IEEE International Conference on Robotics and Automation (ICRA), pp. 3983–3989, 2023

    2023

  34. [34]

    Information theoretic mpc for model-based reinforcement learning,

    G. Williams, N. Wagener, B. Goldfain, P. Drews, J. M. Rehg, B. Boots, and E. A. Theodorou, “Information theoretic mpc for model-based reinforcement learning,” in2017 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2017, pp. 1714–1721. [Online]. Available: https://ieee.org

    2017

  35. [35]

    Terrainnet: Visual modeling of complex terrain for high-speed, off-road navigation

    X. Meng, N. Hatch, A. Lambert, A. Li, N. Wagener, M. Schmittle, J. Lee, W. Yuan, Z. Q. Chen, S. Deng, G. Okopal, D. Fox, B. Boots, and A. Shaban, “Terrainnet: Visual modeling of complex terrain for high-speed, off-road navigation.” inRobotics: Science and Systems,

  36. [36]

    Finding the k shortest loopless paths in a network,

    J. Y . Yen, “Finding the k shortest loopless paths in a network,” Management Science, vol. 17, no. 11, pp. 712–716, 1971

    1971

  37. [2023]

    Available: http://dblp.uni-trier.de/db/conf/rss/rss2023

    [Online]. Available: http://dblp.uni-trier.de/db/conf/rss/rss2023. html#MengHLLWSLYCDOF23

  38. [2025]

    Available: https://arxiv.org/abs/2505.09739

    [Online]. Available: https://arxiv.org/abs/2505.09739

    Pith/arXiv arXiv

  39. [2026]

    Available: https://www.arcgis.com/home/item.html? id=10df2279f9684e4a9f6a7f08febac2a9

    [Online]. Available: https://www.arcgis.com/home/item.html? id=10df2279f9684e4a9f6a7f08febac2a9

This paper was first reviewed by grok-4.5 on July 30, 2026.