Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

MOSU: Autonomous Long-range Robot Navigation with Multi-modal Scene Understanding

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read MOSU claims fusing LiDAR geometry, RGB semantics, and VLM social ranking lifts long-range traversability to 77% on the GND benchmark, a 10% relative improvement over the strongest baseline, with comparable distance to target.

desk verdict A clean but incremental system-integration paper whose headline 10% traversability gain is not yet checkable, because the fusion weights, an ablation, and the promised real-world runs are missing. read the letter →

arxiv 2507.04686 v1 pith:L56RD4I2 submitted 2025-07-07 cs.RO

Jing Liang , Kasun Weerakoon , Daeun Song , Senthurbavan Kirubaharan , Xuesu Xiao , Dinesh Manocha This is my paper
classification cs.RO MSC 68T40
keywords globalnavigationtraversabilityanalysismultimodalperceptionvision-languagemodelssemanticsegmentationtrajectorygenerationsocialoutdoormobilerobots
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

MOSU is a system for long-range outdoor robot navigation whose central claim is that traversability comes from fusing multiple scene-understanding modalities rather than from a single learned planner or an accurate global map. Routing is delegated to public GPS and QGIS services that produce subgoals roughly 50 m apart, and all local decisions are made by generating candidate trajectories and scoring each one with geometric confidence from LiDAR, semantic traversability from RGB segmentation, social and traffic compliance from a vision-language model, and proximity to the GPS subgoal. On the GND benchmark, the fused score selects trajectories with 77% traversability versus 70% for the strongest baseline, while keeping distance-to-target comparable at 73% versus 69%. The paper argues this works because the modalities fail in complementary ways: geometry confuses similar elevations, segmentation fails out of distribution, and language alone is slow, but together they cover each other's gaps. If the claim holds, long-range robot navigation becomes less dependent on maintaining detailed maps and more dependent on rich onboard scene understanding.

What carries the argument

The load-bearing mechanism is the weighted score fusion in Equation 3: $i^* = \arg\max_{i \in [1,N]}\big(\beta_1 c^i_\tau + \beta_2 t^i_\tau + \beta_3 r^i_\tau + \beta_4 g^i_\tau\big)$. Here $c^i_\tau$ is the geometric confidence produced by the CVAE-based trajectory generator from LiDAR point clouds processed by PointCNN plus robot velocities; $t^i_\tau$ is the semantic traversability ratio computed by projecting trajectory waypoints onto Mask2Former image segmentation via Bresenham's algorithm; $r^i_\tau$ is the ranking score from a vision-language model prompted to rank numbered trajectories for social and traffic compliance; and $g^i_\tau$ is the closeness of the trajectory to the current GPS subgoal. The same VLM also returns a velocity mode that caps the Dynamic Window Approach planner at 1 m/s or 0.5 m/s. The routing layer spaces public-routing subgoals about 50 m apart and advances when the robot is within 10 m, so GPS noise never directly steers the robot. The entire argument rests on this particular combination of cheap global guidance and complementary local scores.

What would settle it

Run the full MOSU pipeline on GND campuses not used for any weight selection, with the fusion weights $\beta_1$–$\beta_4$ fixed to the values in the paper once those are reported, and compare traversability to PIVOT; if the advantage drops below the reported 10%, the gain depends on benchmark-specific tuning. A complementary check is to ablate each term in Equation 3 by setting its weight to zero: if removing the semantic term or the VLM ranking term leaves traversability unchanged, the complementarity claim is contradicted.

Watch

Extended reading notes

Core claim

The paper's central claim is that a trajectory-selection fusion of four independent scores achieves the best traversability among compared methods on the GND dataset. The geometric component uses a CVAE-based generator that outputs multiple candidate trajectories from LiDAR point clouds and robot velocities; the semantic component projects those trajectories onto Mask2Former segmentations and computes the fraction of pixels in traversable classes; the VLM component prompts a vision-language model to rank the numbered trajectories according to social and traffic rules; the GPS component scores closeness to the current subgoal. Equation 3 combines these into a single score, and the highest-scoring trajectory goes to a Dynamic Window Approach planner whose maximum speed is set by the VLM's slow or normal mode. The reported result is 77% traversability and 73% distance-to-target, against 70% and 69% for PIVOT, with inference time 2.30 seconds. The claimed discovery is that the complementarity of the modalities, not any single sensor or model, is what yields the improvement.

Load-bearing premise

The load-bearing assumption is that the fixed weights used to combine the four trajectory scores in Equation 3 generalize to new environments; the paper never reports those weights and never tests what happens when one score is removed, so a reader cannot tell whether the 10% gain comes from the fusion idea or from weights tuned to the GND dataset.

Editorial extensions

If this is right

  • Long-range outdoor navigation can be built without a curated global map: public routing plus onboard multimodal trajectory scoring reaches higher traversability than the compared global-navigation baselines.
  • Fusing geometry, semantics, and language covers failure modes that any single modality suffers, so the paper attributes the improvement to complementarity rather than to any one sensor.
  • VLM-based social reasoning can constrain both trajectory choice and robot speed without degrading distance-to-target, since MOSU stays at 73% versus 69% for the strongest baseline.
  • The reported 2.30s inference time is far slower than geometric-only MTG's 0.01s, marking the VLM stage as the practical bottleneck for onboard deployment.

Reading between the lines

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

  • The unreported values of $\beta_1$–$\beta_4$ and the absence of an ablation mean a benchmark-tuned weighting, rather than the fusion concept, could explain part of the 10% gain; a held-out-campus evaluation with frozen weights would settle this.
  • The paper validates the VLM ranking only through qualitative examples and a fixed prompt, not through a social-navigation metric, so the social-compliance contribution is asserted rather than measured.
  • A natural extension the authors do not explore is to learn the fusion weights online or per environment; if the four scores are complementary, adaptive weights should outperform any fixed combination on out-of-distribution terrain.
  • The system's reliance on a single prompt wording and a single vision-language model means the social ranking may be sensitive to both; testing the same fusion across prompts and open-weight models would show how much of the gain is prompt engineering.
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 / 5 minor

Summary. The paper presents MOSU, a three-stage outdoor navigation system consisting of QGIS/GPS-based routing, multimodal trajectory generation, and DWA-based motion planning. The trajectory generation stage combines a CVAE-based geometric trajectory generator (adopted from the authors' MTG work), Mask2Former semantic segmentation for traversability scoring, and a Gemini-based VLM for socially compliant trajectory ranking. The main empirical claim is that MOSU achieves a 10% relative improvement in traversability over the strongest baseline (PIVOT) on the authors' GND dataset (77% vs. 70%) while maintaining comparable distance-to-target (73% vs. 69%). The paper reports results on a 10-campus dataset and provides qualitative examples of geometric, semantic, and VLM-based trajectory evaluation.

Significance. If fully substantiated, the idea of fusing geometric, semantic, and VLM-derived social cues into a single trajectory scoring function would be a useful contribution to outdoor long-range robot navigation. The paper compares against several contemporary baselines (PIVOT, ViNT, NoMaD, MTG, DTG, VL-TGS) and evaluates on a multi-campus dataset, which are strengths. The use of chain-of-thought prompting for VLM-based trajectory ranking is also a reasonable design choice. However, the central quantitative claim cannot currently be verified because the fusion weights and component scores are not reported, no ablation is provided, the distance-to-target metric is ambiguous, and the claimed real-world evaluation is not present in the experiments section.

major comments (4)
  1. [Section 3, Table 1] The traversability and distance-to-target results are reported only as single aggregate numbers over the entire GND dataset, with no per-campus breakdown, no error bars, and no statistical significance tests. Given that the dataset covers 10 campuses with diverse conditions, Table 1 as presented does not support the claim that the 10% improvement over PIVOT is consistent or generalizable; please report per-campus results and variance or confidence intervals, and ideally a paired statistical test across campuses.
  2. [Section 2.2, Eq. (3)] The fusion weights β1–β4 and the component scores ciτ, tiτ, riτ, giτ are never specified or reported. Since the paper attributes the traversability improvement to multimodal fusion, the absence of these values and of an ablation (e.g., removing each modality or setting each β to zero) makes it impossible to determine whether the gain comes from the fusion, from a single modality, or from benchmark-specific tuning. Please report the β values, describe how they were chosen, and include an ablation study over the fusion terms.
  3. [Section 3, 'Distance to Target' metric] The metric is defined as 1 − (dτ − do)/|τ|, but dτ and do are not defined precisely enough to be reproducible, and the expression is not bounded to [0,1]: it can be negative when dτ > do + |τ| and exceeds 1 when dτ < do. Please provide the exact computation per scenario, define dτ and do, and report whether the values in Table 1 are clipped; otherwise the 'comparable distance' claim is not interpretable.
  4. [Abstract vs. Section 3] The abstract states that the system is evaluated 'in real-world on-road environments,' but Section 3 describes only offline evaluation on the GND dataset; no real-world navigation deployment, hardware experiments, or closed-loop results are presented. Please either add such experiments or revise the abstract and claims to match the offline evaluation actually performed.
minor comments (5)
  1. [Author affiliations] The affiliation 'Goerge Mason University' should be corrected to 'George Mason University.'
  2. [References] Reference [3] (Mask2Former) lacks a publication venue and year, and reference [25] (Gemini) lacks full bibliographic information; please complete these entries.
  3. [Section 2.2, VLM prompt] The VLM prompt contains inconsistent goal directions: the first sentence says 'The goal is K meters at [Right Front]' while the later instruction says 'the target is at K meters Front Left'; please align these placeholders.
  4. [Section 2.2, Eq. (1)–(2)] The notation in Eq. (1)–(2) is overloaded, with z_n and N(µ,ν) reused across the derivation; a short algorithm box or pseudocode would make the trajectory-generation process easier to follow.
  5. [Section 2.2, trajectory generator] The trajectory generator is adopted directly from MTG [14]; please state explicitly which parameters are retrained on the GND dataset and which are frozen, since this affects the interpretation of the comparison with MTG in Table 1.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the reported traversability gain is an empirical benchmark result, not a quantity forced by the paper's own equations or by a self-citation chain.

full rationale

The central claim is an empirical comparison on the GND benchmark: MOSU reports 77% traversability versus 70% for PIVOT (Table 1). The derivation chain in Section 2 is a system design: CVAE-based trajectory generation (Eqs. 1-2, adopted from MTG), semantic traversability scoring, VLM ranking, and the weighted fusion rule in Eq. 3. None of these equations is shown by the paper to be equivalent to the evaluation metric by construction. The traversability metric is computed by overlaying generated trajectories on the GND traversability map, while the semantic component uses a separate off-the-shelf segmentation model (Mask2Former) with a shared category vocabulary; sharing a taxonomy is a design choice, not a mathematical reduction. The self-citations to MTG, DTG, VL-TGS, and GND identify published components, baselines, and the benchmark itself, but they do not serve as unverified justifications of the central improvement claim, especially since the table also includes external baselines (PIVOT, ViNT, NoMaD). The unreported fusion weights beta_1-4 and the absence of an ablation are real reproducibility and correctness-risk gaps that could conceal benchmark-specific tuning, but the paper itself provides no quoted equation or construction showing that the 10% gain is forced by fitting those weights. Under the rule that circularity must be exhibited rather than speculated, no concrete circular step can be extracted from the text.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claim rests on a weighted fusion of four scores, with all weights and several thresholds chosen by hand. The method relies on four major domain assumptions about the reliability of the GND labels, MTG confidence, semantic segmentation, and VLM social reasoning. No new physical entities are introduced.

free parameters (5)
  • Fusion weights β1, β2, β3, β4 = not reported
    Weights in Eq. 3 combine geometric confidence, semantic traversability, VLM ranking, and goal distance. Their values are not given and no sensitivity analysis is provided.
  • Candidate trajectory count N = not reported (examples show 5 and 6)
    The number of sampled trajectories affects both geometric coverage and VLM inference time, but no global value is listed.
  • Routing subgoal spacing = 50 m
    Subgoals are spaced approximately 50 m apart to avoid GPS noise, a heuristic choice based on perceptual range.
  • Subgoal proximity threshold = 10 m
    The robot advances to the next subgoal when within 10 m; this threshold is hand-set.
  • Velocity mode limits = normal 1 m/s, slow 0.5 m/s
    Maximum speeds for the two VLM-predicted modes are manually chosen for social compliance.
assumptions (6)
  • domain assumption GND traversability maps are accurate ground truth.
    The evaluation score is the percentage of trajectory waypoints rated traversable by the GND dataset, which is cited from the same research group (Ref. [13]).
  • domain assumption MTG's geometric confidence scores are a valid traversability signal.
    The system adopts the self-cited MTG generator [14] as the geometric component without independent verification on outdoor terrains.
  • domain assumption Mask2Former semantic categories map correctly to traversability for wheeled robots.
    Traversability is defined by five Mask2Former categories; road and sidewalk are treated as navigable, vegetation and building as not, with no calibration for different robot types.
  • domain assumption VLM ranking reflects social and traffic compliance.
    Gemini's textual ranking is used as a scoring signal without quantitative validation of pedestrian distance, traffic rule adherence, or crosswalk usage.
  • domain assumption Public satellite routing provides a safe global path.
    Routing uses Google or OpenRoute services and assumes their suggested paths are traversable by a wheeled robot, which is not verified.
  • ad hoc to paper The four component scores in Eq. 3 are directly comparable.
    Eq. 3 sums geometric confidence, semantic traversability ratio, VLM ranking score, and goal distance score with weights, assuming commensurate scales and additive independence without justification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MOSU: Autonomous Long-range Robot Navigation with Multi-modal Scene Understanding." pith.science (2026). https://pith.science/paper/L56RD4I2

@misc{pith2026250704686,
  author       = {Pith},
  title        = {Pith review of: MOSU: Autonomous Long-range Robot Navigation with Multi-modal Scene Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L56RD4I2}},
  note         = {Machine review of arXiv:2507.04686}
}
read the original abstract

We present MOSU, a novel autonomous long-range navigation system that enhances global navigation for mobile robots through multimodal perception and on-road scene understanding. MOSU addresses the outdoor robot navigation challenge by integrating geometric, semantic, and contextual information to ensure comprehensive scene understanding. The system combines GPS and QGIS map-based routing for high-level global path planning and multi-modal trajectory generation for local navigation refinement. For trajectory generation, MOSU leverages multi-modalities: LiDAR-based geometric data for precise obstacle avoidance, image-based semantic segmentation for traversability assessment, and Vision-Language Models (VLMs) to capture social context and enable the robot to adhere to social norms in complex environments. This multi-modal integration improves scene understanding and enhances traversability, allowing the robot to adapt to diverse outdoor conditions. We evaluate our system in real-world on-road environments and benchmark it on the GND dataset, achieving a 10% improvement in traversability on navigable terrains while maintaining a comparable navigation distance to existing global navigation methods.

Figures

Figures reproduced from arXiv: 2507.04686 by the authors.

Figure 1
Figure 1. Overall System Architecture. MOSU leverages QGIS and GPS to generate long￾range waypoints, serving as high-level guidance for its trajectory generation system. The trajectory generation system provides local trajectories by integrating multimodal perception cues for traversability assessment and leveraging social cues from Vision￾Language Models (VLMs) to ensure social compliance in navigation. This hierarchical app… view at source ↗
Figure 2
Figure 2. Complex Outdoor Scenarios: We evaluate the approach in large-scale, complex outdoor environments with various challenging scenarios, such as (a) narrow spaces, (b) areas with dense off-road vegetation, (c) traffic components (e.g., crosswalks), and (d) social situations involving pedestrians. The metrics in [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Qualitative Evaluation: Each column illustrates one of the three components in our trajectory generation pipeline, (a) CVAE-based multiple trajectory generation, (b) semantic segmentation for traversability analysis, and (c) VLM-based trajectory ranking in the outdoor scenarios. In (a) and (b), lighter colors indicate higher scores, confidence scores in (a) and traversability scores in (b). In (c), the VLM ranks tra… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Geometric and Semantic Analysis: Both geometric trajectory generation and semantic segmentation perform well in most cases, but there are also failure cases, as marked in yellow circles in (c) and (d). When the elevations of the lawn and side￾walk are similar, the geom…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Sem-NaVAE: Semantically-Guided Outdoor Mapless Navigation via Generative Trajectory Priors

    cs.RO 2026-02 conditional novelty 5.0 of 10

    A lightweight CLIPSeg semantic scorer selects among 200 CVAE-generated trajectories, giving 90% success on 120-240 m mapless outdoor routes.

Reference graph

Works this paper leans on

27 extracted references · 24 canonical work pages · cited by 1 Pith paper

  1. [1]

    IEEE Robotics and Automation Letters 7(2) (2022) 3365–3371

    Agishev, R., Petricek, T., Zimmermann, K.: Trajectory optimization using learned robot-terrain interaction model in exploration of large subterranean environments. IEEE Robotics and Automation Letters 7(2) (2022) 3365–3371

    work page 2022

  2. [2]

    In: Seminal graphics: pioneering efforts that shaped the field

    Bresenham, J.E.: Algorithm for computer control of a digital plotter. In: Seminal graphics: pioneering efforts that shaped the field. (1998) 1–6

    work page 1998

  3. [3]

    Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. (2022)

    work page 2022

  4. [4]

    IEEE Intelligent systems 17(2) (2002) 81–83

    Davids, A.: Urban search and rescue robots: from tragedy to technology. IEEE Intelligent systems 17(2) (2002) 81–83

    work page 2002

  5. [5]

    IEEE Robotics & Automation Magazine 4(1) (1997) 23–33

    Fox, D., Burgard, W., Thrun, S.: The dynamic window approach to collision avoidance. IEEE Robotics & Automation Magazine 4(1) (1997) 23–33

    work page 1997

  6. [6]

    In: 2019 IEEE International Conference on Robotics and Biomimetics (ROBIO), IEEE (2019) 1693–1698

    Gao, P., Liu, Z., Wu, Z., Wang, D.: A global path planning algorithm for robots using reinforcement learning. In: 2019 IEEE International Conference on Robotics and Biomimetics (ROBIO), IEEE (2019) 1693–1698

    work page 2019

  7. [7]

    Motion and Operation Planning of Robotic Systems: Background and Practical Approaches (2015) 3–27

    Gasparetto, A., Boscariol, P., Lanzutti, A., Vidoni, R.: Path planning and trajec- tory planning algorithms: A general overview. Motion and Operation Planning of Robotic Systems: Background and Practical Approaches (2015) 3–27

    work page 2015

  8. [8]

    IEEE transactions on pattern analysis and machine intelligence 43(12) (2020) 4338–4364

    Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence 43(12) (2020) 4338–4364

    work page 2020

Show all 27 references
  1. [9]

    Machines 6(3) (2018) 33

    Hoffmann, T., Prause, G.: On the regulatory framework for last-mile delivery robots. Machines 6(3) (2018) 33

    2018

  2. [10]

    IEEE Robotics and Automation Letters (2024)

    Kim, Y., Lee, J.H., Lee, C., Mun, J., Youm, D., Park, J., Hwangbo, J.: Learning semantic traversability with egocentric video and automated annotation strategy. IEEE Robotics and Automation Letters (2024)

    2024

  3. [11]

    In: Pro- ceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Kong, L., Liu, Y., Li, X., Chen, R., Zhang, W., Ren, J., Pan, L., Chen, K., Liu, Z.: Robo3d: Towards robust and reliable 3d perception against corruptions. In: Pro- ceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). (2023) 19994–20006

    2023

  4. [12]

    Advances in neural information processing systems 31 (2018)

    Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: Pointcnn: Convolution on x- transformed points. Advances in neural information processing systems 31 (2018)

    2018

  5. [13]

    In: 2025 IEEE International Conference on Robotics and Automation (ICRA), IEEE (2025)

    Liang*, J., Das*, D., Song*, D., Shuvo, M.N.H., Durrani, M., Taranath, K., Pen- skiy, I., Manocha, D., Xiao, X.: Gnd: Global navigation dataset with multi-modal perception and multi-category traversability in outdoor campus environments. In: 2025 IEEE International Conference ...

    2025

  6. [14]

    In: 2024 IEEE International Conference on Robotics and Automation (ICRA), IEEE (2024) 2396–2402

    Liang, J., Gao, P., Xiao, X., Sathyamoorthy, A.J., Elnoor, M., Lin, M.C., Manocha, D.: Mtg: Mapless trajectory generator with traversability coverage for outdoor navigation. In: 2024 IEEE International Conference on Robotics and Automation (ICRA), IEEE (2024) 2396–2402

    2024

  7. [15]

    In: 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), IEEE (2024) 5340–5347

    Liang, J., Payandeh, A., Song, D., Xiao, X., Manocha, D.: Dtg: Diffusion-based tra- jectory generation for mapless global navigation. In: 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), IEEE (2024) 5340–5347

    2024

  8. [16]

    arXiv preprint arXiv:2402.07872 (2024)

    Nasiriany, S., Xia, F., Yu, W., Xiao, T., Liang, J., Dasgupta, I., Xie, A., Driess, D., Wahid, A., Xu, Z., et al.: Pivot: Iterative visual prompting elicits actionable knowledge for vlms. arXiv preprint arXiv:2402.07872 (2024)

    2024 arXiv

  9. [17]

    Ocean Engi- neering 251 (2022) 111010 10 Jing Liang et al

    Ozturk, U., Akdaug, M., Ayabakan, T.: A review of path planning algorithms in maritime autonomous surface ships: Navigation safety perspective. Ocean Engi- neering 251 (2022) 111010 10 Jing Liang et al

    2022

  10. [18]

    In: 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), IEEE (2024) 13837–13844

    Sathyamoorthy, A.J., Weerakoon, K., Elnoor, M., Zore, A., Ichter, B., Xia, F., Tan, J., Yu, W., Manocha, D.: Convoi: Context-aware navigation using vision language models in outdoor and indoor environments. In: 2024 IEEE/RSJ International Conference on Intelligent Robots and S...

    2024

  11. [19]

    arXiv preprint arXiv:2306.14846 (2023)

    Shah, D., Sridhar, A., Dashora, N., Stachowicz, K., Black, K., Hirose, N., Levine, S.: Vint: A foundation model for visual navigation. arXiv preprint arXiv:2306.14846 (2023)

    2023 arXiv

  12. [20]

    Information Fusion (2024) 102601

    Sohail, S.S., Himeur, Y., Kheddar, H., Amira, A., Fadli, F., Atalla, S., Copiaco, A., Mansoor, W.: Advancing 3d point cloud understanding through deep transfer learning: A comprehensive survey. Information Fusion (2024) 102601

    2024

  13. [21]

    Advances in neural information processing systems 28 (2015)

    Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. Advances in neural information processing systems 28 (2015)

    2015

  14. [22]

    IEEE Robotics and Automation Letters 10(1) (2025) 508–515

    Song, D., Liang, J., Payandeh, A., Raj, A.H., Xiao, X., Manocha, D.: Vlm-social- nav: Socially aware robot navigation through scoring using vision-language models. IEEE Robotics and Automation Letters 10(1) (2025) 508–515

    2025

  15. [23]

    IEEE Robotics and Automation Letters 10(6) (2025) 5791–5798

    Song, D., Liang, J., Xiao, X., Manocha, D.: Vl-tgs: Trajectory generation and selection using vision language models in mapless outdoor environments. IEEE Robotics and Automation Letters 10(6) (2025) 5791–5798

    2025

  16. [24]

    In: 2024 IEEE International Conference on Robotics and Automation (ICRA), IEEE (2024) 63–70

    Sridhar, A., Shah, D., Glossop, C., Levine, S.: Nomad: Goal masked diffusion policies for navigation and exploration. In: 2024 IEEE International Conference on Robotics and Automation (ICRA), IEEE (2024) 63–70

    2024

  17. [25]

    Team, G.: Gemini: A family of highly capable multimodal models (2024)

    2024

  18. [26]

    In: Proceedings of the 28th international technical meeting of the satellite division of the institute of navigation (ION GNSS+ 2015)

    Van Diggelen, F., Enge, P.: The world’s first gps mooc and worldwide laboratory using smartphones. In: Proceedings of the 28th international technical meeting of the satellite division of the institute of navigation (ION GNSS+ 2015). (2015) 361–369

    2015

  19. [27]

    IEEE access 8 (2020) 58443– 58469

    Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driv- ing: Common practices and emerging technologies. IEEE access 8 (2020) 58443– 58469

    2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.