Pith. sign in

REVIEW 3 major objections 5 minor 27 references

A frozen multimodal language model — with only its language layers adapted and no dedicated vision encoder or regression head — can issue metric waypoint trajectories as discrete tokens and reliably reach specified goal coordinates in four

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

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

T0 review

2026-08-02 08:10 UTC pith:SCVYOYFZ

load-bearing objection Frozen-vision discrete-token navigation is a real result; the absolute stopping-error claim is only as good as the unquantified SLAM frame. the 3 major comments →

arxiv 2607.06882 v2 pith:SCVYOYFZ submitted 2026-07-08 cs.RO cs.AI

GemNav: Discrete-Token Visual Robot Navigation using a Multimodal Large Language Model

Peter Bohm , Saimunur Rahman , Abdelwahed Khamis , Sagun Man Singh Shrestha , Chris McCool , Peyman Moghadam This is my paper
classification cs.RO cs.AI
keywords visual navigationmultimodal large language modeldiscrete-token action representationlow-rank adaptationfrozen vision encoderzero-shot transferwaypoint predictiondata-efficient robot learning
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.

GemNav asks whether the standard recipe for vision-language-action navigation — a dedicated visual encoder, a bespoke action head, and thousands of hours of training data — is necessary. The paper claims it is not: a frozen multimodal LLM, adapted only with low-rank updates to its language tower, can turn a current camera view and a goal specification into an eight-waypoint trajectory expressed as discrete tokens, plus a stop signal. Trained on a single open 8.7-hour teleoperated corpus, the policy transfers zero-shot to four physically distinct unseen environments and stops within 0.25–0.42 m of the goal on all 20 pose-conditioned real-world trials. A soft-decoded auxiliary loss recovers the metric ordering that pure token cross-entropy discards, and the authors report that adding short image histories improves offline metrics but not closed-loop robot performance. If correct, this suggests action representation — not further visual scaling — may be the decisive design axis for adapting MLLMs to navigation.

Core claim

The central claim is that waypoint navigation can be reformulated as conditional token generation in a frozen multimodal LLM. A 2D goal can be given as an image, a local coordinate, or both; the model outputs either an 18-token sequence encoding eight body-frame waypoints via 64 value bins spanning ±15 m, or a categorical stop token (<goal reached> / <goal unreachable>). LoRA adaptation is applied only to the language tower's linear layers; the vision tower stays completely frozen and no continuous regression head is added. A soft-decoded expectation over bin centers, supervised by an auxiliary MSE term, restores metric structure to the discrete predictions and cuts validation displacement e

What carries the argument

The load-bearing object is the discrete value-bin token vocabulary: 64 reserved tokens uniformly quantize the [-15 m, 15 m] range per axis, so every waypoint coordinate, every goal coordinate, and the stop tokens all live in one vocabulary produced by the stock language-model head. The trajectory target is an 18-token sequence wrapped by <bot traj>/<eot traj>; stop decisions are the same wrapper with a single categorical token. LoRA (rank 32, alpha 16) adapts only language-tower linear layers. The soft-decoded auxiliary loss computes a softmax expectation over bin centers and applies an MSE against true coordinates, which gives the model ordinal feedback without adding parameters or breaking

Load-bearing premise

The 0.25–0.42 m stopping distances are measured in the coordinate frame of the robot's online continuous-time SLAM system, and the paper gives no drift or bias estimate, so the reported meters are not anchored to independent ground truth.

What would settle it

Repeat the four-environment deployment while logging an independent localization source (for example, RTK GPS or motion capture) at the goal poses. If the difference between the SLAM-reported goal coordinate and the independent measurement is comparable to or larger than the reported 0.25–0.42 m stopping error, the claimed metric accuracy is an artefact of a consistent training/deployment frame rather than genuine spatial grounding.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A navigation policy can be trained from a single open corpus of about 8.7 hours—roughly three orders of magnitude less data than typical cross-embodiment navigation sets—and still succeed zero-shot in four unseen environments.
  • The standard recipe of a dedicated visual encoder plus a continuous regression head is not required for short-to-medium waypoint navigation; frozen pretrained visual features are sufficient.
  • Expressing metric waypoints, goal coordinates, and stop decisions through one discrete token vocabulary lets the model use a uniform interface for conditioning and generation, which the paper argues is what makes pose-plus-image conditioning work.
  • The soft-decoded auxiliary regression loss reduces validation displacement error by 23% over pure cross-entropy, so ordinal structure can be recovered without adding parameters.
  • Visual history (two or four frames) improves offline ADE but does not improve real-robot stopping, so temporal context is not an automatic win once pretrained vision features are in place.

Where Pith is reading between the lines

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

  • An independent localization check (for example, RTK GPS or motion capture) would settle whether the 0.25–0.42 m stopping distances are true world-frame errors or artefacts of using the same SLAM frame for training and deployment.
  • If the token-interface recipe generalizes, the same frozen MLLM could accept language goals or planner critiques as additional tokens on the conditioning side, making navigation policies easier to steer without retraining vision or adding heads.
  • The negative history-depth result suggests offline ADE is a weak proxy for closed-loop behaviour; rollout-based validation or explicit modelling of past motion may be needed before temporal context can help.
  • Because image-only stopping was unreliable, a practical extension would be to make the policy express in-place rotation or to fuse richer goal descriptors, removing the dependence on a metric coordinate.

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 / 5 minor

Summary. The paper presents GemNav, a waypoint-navigation policy that adapts a frozen Gemma-4-E2B-it multimodal LLM with LoRA applied only to the language tower, using a discrete token vocabulary for eight body-frame waypoints and categorical stop signals, plus a soft-decoded auxiliary loss to recover metric structure. The model is trained on roughly 8.7 hours of SCAND image/pose data and evaluated offline on SCAND and a newly collected TokenWalker dataset, and on a Spot robot in four unseen environments. The paper reports 20/20 pose-goal successes with final displacements of 0.25–0.42 m, while noting that image-only goal reaching is unreliable. Ablations address decode mode, history depth, goal sampling, LoRA rank, and crop mode.

Significance. If the real-robot results are trustworthy, this is a significant data-efficiency result: it suggests that a frozen MLLM's multimodal grounding, combined with a discrete output interface and a light LoRA adapter, can support deployable short-to-medium-horizon navigation with orders of magnitude less training data than cross-embodiment VLAs. The tokenization scheme and the soft-decoded auxiliary loss are clean and well motivated, and the ablation in Appendix E.1 supporting a 23% ADE reduction is a concrete, internally consistent contribution. The negative result on image-history depth is also useful, with an explicit matched-window control in Appendix I. The main weakness is that the headline deployment metric is not anchored to an independent localization source.

major comments (3)
  1. [§4.3, §4.6, §5.2, Appendix L] The central real-robot claim—stopping within 0.25–0.42 m over 20 trials—is measured entirely in the frame of the online continuous-time LiDAR-inertial SLAM [25]. Goal coordinates are supplied in that same frame and final displacement is computed from it. Appendix L says the estimator maintains 'consistency' with the corpus pose pipeline, but it gives no drift or accuracy quantification, and the SCAND pose source is not specified. On 30–34 m runs, even a well-tuned lidar-inertial SLAM can accumulate drift of tens of centimeters, which would make the reported stopping errors comparable to the measurement noise floor. This is load-bearing for the 20/20 success claim and for the comparison with OmniVLA's final displacement. Please provide a quantitative characterization of the online SLAM accuracy (e.g., loop-closure residuals, comparison with RTK/GPS or a surveyed ground-truth map, or repea
  2. [Abstract, §5.2, Table 16, §7] The abstract states that the policy 'stops within 0.25–0.42 m of the goal across 20 real-world trials' without specifying that these are exclusively pose-goal trials. The image-goal (ego) modality succeeds only 1/5 in the Carpark (Table 16), is not evaluated in the other environments, and §7 states that auto-stop on a pure image goal is unreliable. As written, the abstract overstates the coverage of the success claim. Please qualify the abstract and the results summary to make clear that the 20/20 result holds for pose-conditioned goals only, and that image-goal navigation remains an open limitation.
  3. [§6, Appendix I, Table 12] The conclusion that visual history 'yields no robot benefit' is too strong given Table 12. For the ego+pose modality, the h=2 checkpoint achieves 5/5 autonomous stops while the h=0 checkpoint achieves 4/5, and the h=2 checkpoint also has smaller final displacement on ego+pose (0.30 m vs 0.33 m). The h=2 policy is slower, so the trade-off is real, but the statement as written in the abstract and Section 6 is contradicted by the paper's own table. Please soften or qualify this claim.
minor comments (5)
  1. [§4.4] The term 'dLoRA' is used without definition. If it is simply LoRA, please use 'LoRA' consistently.
  2. [Table 2] The Warehouse row uses n=4 for OmniVLA but n=5 for other cells. The reason for this exception should be stated in the table caption or text.
  3. [Appendix D] Several entries in Table 4 and the surrounding text are marked 'pending re-eval' or measured against an earlier, lagged-position TokenWalker manifest. Such provisional values should either be updated to final numbers or removed, since they do not yet support the crop-mode comparisons.
  4. [Appendix L] The phrase 'maintains consistency with the pose-estimation pipeline' is vague. If the authors cannot add quantitative SLAM error analysis, at least explain what 'consistency' means operationally (same estimator configuration? same coordinate convention? calibration?).
  5. [§4.6] The success metric is defined as an autonomous <goal_reached> stop, but for OmniVLA, which has no stop signal, the paper uses stopping pose distance. This asymmetry is acknowledged, but it is worth stating explicitly in the main text that the comparison is between different termination criteria, not identical tasks.

Circularity Check

0 steps flagged

No significant circularity: tokenization, training, and held-out evaluation are self-contained; the only author-overlap citation (Wildcat SLAM) is localization infrastructure, not a load-bearing derivation step.

full rationale

The core derivation chain is not circular. Eq. (1) defines a fixed 64-bin quantization over [-15,15]; Eq. (2) maps ground-truth 8-waypoint trajectories to 18-token sequences; Eqs. (3)-(4) add a soft-decoded MSE auxiliary loss against those same ground-truth waypoints. This is ordinary supervised sequence prediction, and the evaluation is held out: a 10% SCAND trajectory-disjoint split and the entirely withheld TokenWalker corpus, with deployment in four unseen environments. No tokenization constant, loss weight, or architecture choice is fit to the test results; the claimed 0.25-0.42 m final displacements come from closed-loop trials, not from the training objective. The one author-overlap citation is [25], Wildcat SLAM (P. Moghadam is a GemNav author), used only to supply body-frame poses at deployment and for corpus poses. That is externally published localization infrastructure, not an imported uniqueness theorem or a fitted parameter, so it is not a circular step. Appendix L's assertion that Wildcat 'maintains consistency with the pose-estimation pipeline used to record the corpus' is a measurement premise; if the SLAM frame drifts, the reported meters are a validity/correctness concern, not a reduction of the policy's prediction to its input. The paper also explicitly flags its own limitations (Section 7) and an OOD manifest-mismatch caveat (Appendix E.3), consistent with a non-circular evaluation. Score 2 reflects the minor self-citation and the unquantified localization premise, neither of which is structurally circular.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 1 invented entities

The central result rests on explicitly stated design choices (K=64, B=15m, λ_aux=0.1, LoRA rank, history depth) and on three domain assumptions: frozen-vision sufficiency, SCAND representativeness, and SLAM accuracy. SLAM accuracy is the least supported; the paper provides no error quantification. The token vocabulary is a new representational entity confined to this paper.

free parameters (6)
  • K = 64 value-bin tokens = 64
    Number of uniform bins on [-15,15] m; chosen so half-bin error (0.234m/axis) fits the 0.5m goal-arrival radius. Affects achievable final displacement; hand-set, not learned.
  • B = 15m value-bin range = 15 m
    Set with ~7% margin above max training goal arc-length (14m); hand-set; determines bin width along with K.
  • Auxiliary-loss weight λ_aux = 0.1
    Fixed weight on soft-decoded regression loss; no ablation on λ reported; directly changes the training objective.
  • LoRA rank r and scale α = r=32, α=16
    Adapter capacity; ablation shows r=64 worse, attributed to α/r scaling; hand-tuned.
  • History depth h = 0 (deployed)
    Deployment choice after h=2,4 reduce closed-loop success; offline ADE improves with h, so this is a design decision dependent on deployment evaluation.
  • Goal sampler distance range = Uniform(0.5, 14) m
    Training target goals sampled by arc-length; replaces frame-aligned sampler after deployment failure; hand-chosen distribution.
axioms (5)
  • domain assumption Gemma-4-E2B-it's frozen vision tower provides sufficient visual features for waypoint navigation
    Core hypothesis of the paper (Section 1, 3.3); supported by experiments but not independently proven.
  • domain assumption Online SLAM (Wildcat) body-frame poses are accurate enough for metric goal coordinates and final-displacement measurements
    Used in Sections 4.3 and L; no error/drift quantification provided; success metrics are internal to the SLAM frame.
  • domain assumption SCAND's 8.7h corpus generalizes to unseen industrial/warehouse/outdoor environments
    The zero-shot claim rests on this (Section 4.1, Appendix O); tested but with n=5 trials per environment.
  • standard math LoRA adaptation and the soft-decoded auxiliary loss train effectively via standard backprop
    Assumes gradients through softmax and LoRA updates work as in [23] and Eq. 3-4.
  • domain assumption An 8-waypoint body-frame (x,y) sequence plus external collision avoidance is a sufficient action interface
    Filters in Appendix K exclude goals requiring rotation; policy cannot express in-place rotation (Limitations).
invented entities (1)
  • Navigation token vocabulary (5 role tokens + 64 value-bin tokens) no independent evidence
    purpose: Represent waypoints, goal coordinates, and stop signals (<bot traj>, <eot traj>, <goal reached>, <goal unreachable>, <goal coord>, and 64 value bins) through the LM head
    Tokens are allocated from Gemma-4's reserved <unusedN> range (Section 3.2); their function is demonstrated within this paper only, with no external benchmark isolating their contribution.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of GemNav: Discrete-Token Visual Robot Navigation using a Multimodal Large Language Model." pith.science (2026). https://pith.science/paper/SCVYOYFZ

@misc{pith2026260706882,
  author       = {Pith},
  title        = {Pith review of: GemNav: Discrete-Token Visual Robot Navigation using a Multimodal Large Language Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SCVYOYFZ}},
  note         = {Machine review of arXiv:2607.06882}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Visual navigation policies built on large pretrained models have so far followed a common recipe: a dedicated visual encoder, a bespoke action head, and training on thousands of hours of cross-embodiment datasets. We ask whether this recipe is necessary. In this paper, we introduce GemNav, a visual robot navigation policy that adapts a frozen Multimodal Large Language Model (MLLM) for short-to-medium horizon waypoint navigation using Low-Rank Adaptation (LoRA) on the language tower alone, with no auxiliary visual encoder and no continuous regression head. Waypoints and categorical navigation signals share a single discrete token vocabulary generated by the language-model head, and a soft-decoded auxiliary loss recovers the metric structure that pure cross-entropy training discards. On a single 8.7-hour open corpus, roughly three orders of magnitude smaller than competing training sets, the policy transfers zero-shot to four physically distinct unseen environments and stops within 0.25-0.42m of the goal across 20 real-world trials covering an open carpark, an obstacle carpark, a long outdoor chemical yard, and an indoor warehouse. Conditioning on short image histories improves offline metrics but yields no robot benefit, pointing to a ceiling on what temporal context adds once pretrained vision features are in place. These results indicate that discrete-token adaptation of frozen MLLMs can provide a data-efficient, deployable alternative for foundation model robot navigation.

Figures

Figures reproduced from arXiv: 2607.06882 by Abdelwahed Khamis, Chris McCool, Peter Bohm, Peyman Moghadam, Sagun Man Singh Shrestha, Saimunur Rahman.

Figure 1
Figure 1. Figure 1: GemNav architecture. Inputs: current camera view, optional goal image and/or 2D goal coordinate, and (grayed, not yet implemented) a language prompt. Images pass through Gemma 4’s frozen vision tower; 2D goal coordinates reuse the trajectory value-bin alphabet, prefixed by <goal coord>. The language model is adapted via LoRA on its linear layers (r=32, α=16). Outputs are an 8-waypoint trajectory (<bot traj… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative single-obstacle warehouse run. Top: scene with start, obstacle, and goal. Middle: time-lapse of GemNav (left) reaching the goal and OmniVLA (right) walking past it to the far wall. Bottom: overlaid trajectories, GemNav routing around the obstacle and stopping on the goal (success), OmniVLA holding its heading past the goal (failure) Because OmniVLA already fails in the presence of obstacles, th… view at source ↗
Figure 3
Figure 3. Figure 3: Medium-horizon indoor warehouse run ( 32 m, unseen scene). Top: time-lapse of GemNav on Spot traversing the course from start (right) to goal (left). Despite dense clutter (bins, cones, workbenches, machinery, and a seated bystander, highlighted in blue) and a goal not visible from the start, the policy reaches the target successfully. Bottom: executed path inferred from visual inspection. Assistant turn (… view at source ↗

Review history (2 revisions) →

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

27 extracted references · 2 linked inside Pith

  1. [1]

    Firoozi, J

    R. Firoozi, J. Tucker, S. Tian, A. Majumdar, J. Sun, W. Liu, Y . Zhu, S. Song, A. Kapoor, K. Hausman, et al. Foundation models in robotics: Applications, challenges, and the future. The International Journal of Robotics Research, 44(5):701–739, 2025

    2025

  2. [2]

    D. Shah, A. Sridhar, N. Dashora, K. Stachowicz, K. Black, N. Hirose, and S. Levine. Vint: A foundation model for visual navigation. In7th Annual Conference on Robot Learning (CoRL), pages 711–733, 2023

    2023

  3. [3]

    Sridhar, D

    A. Sridhar, D. Shah, C. Glossop, and S. Levine. Nomad: Goal masked diffusion policies for navigation and exploration. In2024 IEEE International Conference on Robotics and Automa- tion (ICRA), pages 63–70. IEEE, 2024

    2024

  4. [4]

    Zhang, A

    J. Zhang, A. Li, Y . Qi, M. Li, J. Liu, S. Wang, H. Liu, G. Zhou, Y . Wu, X. Li, Y . Fan, W. Li, Z. Chen, F. Gao, Q. Wu, Z. Zhang, and H. Wang. Embodied navigation foundation model. arXiv preprint arXiv:2509.12129, 2025

    arXiv 2025

  5. [5]

    Hirose, C

    N. Hirose, C. Glossop, D. Shah, and S. Levine. Omnivla: An omni-modal vision-language- action model for robot navigation.IEEE International Conference on Robotics and Automa- tion, 2026

    2026

  6. [6]

    Doshi, H

    R. Doshi, H. R. Walke, O. Mees, S. Dasari, and S. Levine. Scaling cross-embodied learning: One policy for manipulation, navigation, locomotion and aviation. InConference on Robot Learning, pages 496–512. PMLR, 2025

    2025

  7. [7]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. P. Foster, P. R. Sanketi, Q. Vuong, et al. Openvla: An open-source vision-language-action model. InConference on Robot Learning, pages 2679–2713. PMLR, 2025

    2025

  8. [8]

    Zhang, X

    J. Zhang, X. Chen, Q. Wang, M. Li, Y . Guo, Y . Hu, J. Zhang, S. Bai, J. Lin, and J. Chen. Vlm4vla: Revisiting vision-language-models in vision-language-action models.International Conference on Learning Representations, 2026

    2026

  9. [9]

    Dey, J.-N

    S. Dey, J.-N. Zaech, N. Nikolov, L. Van Gool, and D. P. Paudel. Revla: Reverting visual domain limitation of robotic foundation models. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 8679–8686, 2025

    2025

  10. [10]

    Gemma 4 model.https://ai.google.dev/gemma/docs/core/model_ card_4, 2026

    Gemma Team. Gemma 4 model.https://ai.google.dev/gemma/docs/core/model_ card_4, 2026. Google DeepMind

    2026

  11. [11]

    Zitkovich, T

    B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahid, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In Conference on Robot Learning, pages 2165–2183. PMLR, 2023

    2023

  12. [12]

    W. Yuan, J. Duan, V . Blukis, W. Pumacay, R. Krishna, A. Murali, A. Mousavian, and D. Fox. Robopoint: A vision-language model for spatial affordance prediction in robotics. In8th Annual Conference on Robot Learning

  13. [13]

    Driess, F

    D. Driess, F. Xia, M. S. Sajjadi, C. Lynch, A. Chowdhery, B. Ichter, A. Wahid, J. Tompson, Q. Vuong, T. Yu, et al. Palm-e: an embodied multimodal language model. InProceedings of the 40th International Conference on Machine Learning, pages 8469–8488, 2023

    2023

  14. [14]

    D. Shah, A. Sridhar, A. Bhorkar, N. Hirose, and S. Levine. Gnm: A general navigation model to drive any robot. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7226–7233. IEEE, 2023

    2023

  15. [15]

    S. Zeng, D. Qi, X. Chang, F. Xiong, S. Xie, X. Wu, S. Liang, M. Xu, X. Wei, and N. Guo. Janusvln: Decoupling semantics and spatiality with dual implicit memory for vision-language navigation.International Conference on Learning Representations, 2025. 9

    2025

  16. [16]

    Hirose, C

    N. Hirose, C. Glossop, A. Sridhar, O. Mees, and S. Levine. Lelan: Learning a language- conditioned navigation policy from in-the-wild video. In8th Annual Conference on Robot Learning

  17. [17]

    X. Liu, J. Li, Y . Jiang, N. Sujay, Z. Yang, J. Zhang, J. Abanes, J. Zhang, and C. Feng. City- walker: Learning embodied urban navigation from web-scale videos. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 6875–6885, 2025

    2025

  18. [18]

    Glossop, W

    C. Glossop, W. Chen, A. Bhorkar, D. Shah, and S. Levine. Cast: Counterfactual labels improve instruction following in vision-language-action models.arXiv preprint arXiv:2508.13446, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    Karnan, A

    H. Karnan, A. Nair, X. Xiao, G. Warnell, S. Pirk, A. Toshev, J. Hart, J. Biswas, and P. Stone. Socially compliant navigation dataset (scand): A large-scale dataset of demonstrations for social navigation.IEEE Robotics and Automation Letters, 7(4):11807–11814, 2022

    2022

  20. [20]

    A. J. Hancock, X. Wu, L. Zha, O. Russakovsky, and A. Majumdar. Actions as language: Fine- tuning vlms into vlas without catastrophic forgetting.International Conference on Learning Representations, 2026

    2026

  21. [21]

    Z. Hou, T. Zhang, Y . Xiong, H. Duan, H. Pu, R. Tong, C. Zhao, X. Zhu, Y . Qiao, J. Dai, et al. Dita: Scaling diffusion transformer for generalist vision-language-action policy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7686– 7697, 2025

    2025

  22. [22]

    P. Li, Y . Chen, H. Wu, X. Ma, X. Wu, Y . Huang, L. Wang, T. Kong, and T. Tan. Bridgevla: Input-output alignment for efficient 3d manipulation learning with vision-language models. Advances in Neural Information Processing Systems, 38:63635–63673, 2026

    2026

  23. [23]

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chen, et al. Lora: Low-rank adaptation of large language models.International Conference on Learning Repre- sentations, 1(2):3, 2022

    2022

  24. [24]

    Karamcheti, S

    S. Karamcheti, S. Nair, A. Balakrishna, P. Liang, T. Kollar, and D. Sadigh. Prismatic vlms: Investigating the design space of visually-conditioned language models. InForty-first Interna- tional Conference on Machine Learning, 2024

    2024

  25. [25]

    Ramezani, K

    M. Ramezani, K. Khosoussi, G. Catt, P. Moghadam, J. Williams, P. Borges, F. Pauling, and N. Kottege. Wildcat: Online continuous-time 3d lidar-inertial slam.arXiv preprint arXiv:2205.12595, 2022

    Pith/arXiv arXiv 2022

  26. [26]

    Alahi, K

    A. Alahi, K. Goel, V . Ramanathan, A. Robicquet, L. Fei-Fei, and S. Savarese. Social lstm: Human trajectory prediction in crowded spaces. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 961–971, 2016

    2016

  27. [27]

    drives past and ignores

    T. Salzmann, B. Ivanovic, P. Chakravarty, and M. Pavone. Trajectron++: Dynamically-feasible trajectory forecasting with heterogeneous data. InEuropean conference on computer vision, pages 683–700. Springer, 2020. 10 Supplementary Material for GemNav: Discrete-Token Visual Robot Navigation using a Multimodal Large Language Model A TokenWalker Dataset (TW) ...

    2020

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.