Pith. sign in

REVIEW 3 major objections 8 minor 27 references

Frozen vision model steers a robot with 8.7 hours of data

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 · glm-5.2

2026-07-09 23:36 UTC pith:SCVYOYFZ

load-bearing objection Frozen MLLM + discrete action tokens + 8.7 hours of data = real robot navigation that works. The OmniVLA comparison is the weak point. the 3 major comments →

arxiv 2607.06882 v1 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 tokenizationlow-rank adaptationrobot learningdata-efficient learningwaypoint prediction
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 paper argues that the dominant recipe in foundation-model robot navigation — a dedicated visual encoder trained on thousands of hours of cross-embodiment data, paired with a continuous regression head — is not necessary. The authors adapt a frozen multimodal large language model (Gemma 4) by applying low-rank adaptation only to its language tower, leaving the vision encoder untouched and adding no separate action head. Navigation waypoints and stop signals are expressed as discrete tokens from the model's own language-model head, and a soft-decoded auxiliary loss reintroduces metric structure that pure cross-entropy discards. Trained on a single 8.7-hour open dataset, the resulting policy reaches goals within 0.25–0.42 meters across four unseen real-world environments, including obstacle courses, an outdoor chemical yard, and an indoor warehouse. A secondary finding is that adding short visual histories improves offline trajectory metrics but degrades closed-loop robot performance, suggesting diminishing returns for temporal context once pretrained vision features are in place.

Core claim

The central claim is that a frozen pretrained MLLM's visual representations are already sufficient for short-to-medium-horizon waypoint navigation, provided the action representation is rethought. By quantizing continuous waypoints into discrete value-bin tokens and generating them through the stock language-model head, the authors show that metric navigation and categorical stop decisions can share a single token interface without any auxiliary visual encoder or regression head. The soft-decoded auxiliary loss — which computes a weighted expectation over bin centers to recover sub-bin precision — reduces validation displacement error by 23% over pure cross-entropy while preserving the token

What carries the argument

The mechanism is a discrete value-bin tokenizer: each waypoint axis is independently quantized into one of 64 uniform bins spanning ±15 meters, producing an 18-token trajectory sequence (two role tokens plus 16 coordinate tokens). A soft-decoded auxiliary loss decodes the bin logits into a continuous prediction via a weighted sum over bin centers, penalizing distant bin errors more heavily than adjacent ones. Goal coordinates use the same value-bin tokens as inputs, eliminating any representational gap between conditioning and generation. The entire policy is adapted via LoRA (rank 32, alpha 16) on all language-tower linear layers, with the vision tower frozen and no additional parameters.

Load-bearing premise

The claim that frozen pretrained vision features are sufficient for navigation rests on deployment environments whose visual characteristics overlap substantially with the training data. All four test sites and the training corpus cover structured outdoor and indoor spaces with paved surfaces, industrial settings, or parking areas. Whether the frozen vision tower's features would hold up in fundamentally different visual domains is not tested.

What would settle it

If a frozen MLLM's pretrained vision features were truly sufficient for navigation, then deploying the same LoRA-adapted policy in a visually distinct domain — say, dense forest trails or dynamic crowded indoor spaces — should still produce goal-reaching behavior. If it fails, the sufficiency claim is specific to visually structured environments similar to the training distribution.

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

If this is right

  • If frozen MLLM vision features suffice for navigation, the field's trajectory of scaling visual encoders on robot data may be addressing a bottleneck that pretrained representations already solve.
  • The discrete-token action interface could generalize beyond navigation to other continuous-control tasks where ordinal structure matters, since the soft-decoding auxiliary loss is architecture-agnostic.
  • The finding that offline metrics improve with visual history but closed-loop performance degrades challenges the common assumption that better logged-trajectory scores predict better deployment.
  • Distance-based goal sampling — selecting goals by metric arc-length rather than fixed frame offsets — may be a critical but underappreciated design choice for deployable navigation policies.

Where Pith is reading between the lines

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

  • The claim that visual scaling is unnecessary rests on deployment environments that share visual characteristics with the training data (paved surfaces, industrial settings, parking areas). Whether frozen MLLM features generalize to fundamentally different visual domains — dense vegetation, dynamic crowds, extreme lighting — remains untested.
  • The 2D waypoint output cannot express in-place rotation, and goals behind the robot are filtered during training. This constrains the policy to forward-facing navigation and may limit applicability in scenarios requiring omnidirectional planning.
  • The time-lagged trajectory effect observed with visual history suggests that the policy may be learning trajectory priors rather than reactive visual grounding, which could interact poorly with dynamic environments.
  • If the soft-decoded auxiliary loss principle generalizes, it could apply to any task where discrete tokenization of continuous values loses ordinal information, potentially including manipulation force prediction or grasp pose regression.

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. GemNav adapts a frozen Gemma-4 MLLM for short-to-medium-horizon waypoint navigation by applying LoRA solely to the language tower, representing waypoints and stop signals as discrete tokens from the model's native LM head. A soft-decoded auxiliary loss reintroduces metric structure over value bins. The system is trained on the 8.7-hour SCAND dataset and evaluated in four unseen real-world environments (open carpark, obstacle carpark, outdoor chemical yard, indoor warehouse) on a Boston Dynamics Spot, achieving 20/20 pose-goal successes with 0.25-0.42m stopping precision. The paper also reports a negative result: visual history improves offline ADE but degrades closed-loop deployment, and introduces a distance-based goal sampler that proves critical for deployability.

Significance. The paper addresses a timely and well-posed question: whether the frozen visual representations of a modern MLLM are sufficient for deployable robot navigation, or whether the field's standard recipe (dedicated visual encoder + continuous regression head + massive cross-embodiment datasets) is necessary. The discrete-token action representation with a soft-decoded auxiliary loss (Section 3.2, Eqs. 3-4) is a clean, parameter-free contribution that reduces validation ADE by 23% (Appendix E.1). The distance-based goal sampler ablation (Section 6, Appendix F) provides a falsifiable and practically useful finding: in-distribution ADE alone does not predict deployability. The history-depth negative result (Section 6, Appendix I) is valuable and honestly reported. Real-robot deployment across four environments with consistent stopping precision is a substantive empirical contribution. The data-efficiency claim (8.7 hours vs. thousands) is attention-grabbing but its strength depends on the fairness of the OmniVLA comparison, which has issues discussed below.

major comments (3)
  1. The OmniVLA baseline comparison is structurally compromised by a metric mismatch that undermines the headline 20/20 vs 0/19 success-rate claim. Section 4.4 states OmniVLA 'does not autonomously generate stop signals,' so its SR is 0 by construction across all environments. The primary headline metric therefore compares a model that can stop with one that cannot, which is uninformative as a success-rate comparison. The paper does report nearest-approach distances for OmniVLA (Table 2, e.g., 9.54m on Carpark, 12.99m on Chemical Yard), which are the informative numbers, but these are buried in parentheses and the narrative frames the SR tally as the central comparison. The paper should either (a) reframe the OmniVLA comparison around nearest-approach and path-quality metrics rather than SR, or (b) implement a heuristic stop rule for OmniVLA (e.g., stop when displacement to goal falls below
  2. The OmniVLA baseline shows a 15x discrepancy in nearest-approach quality across environments that is unexplained and suggests possible configuration issues. In Table 2 (Carpark, pose goal), OmniVLA's nearest approach is 9.54m on a ~10-12m course — a navigation failure. But in Table 3 (Warehouse, Open, ego+pose), its nearest approach is 0.65m with 2.87m final displacement — it reaches the goal vicinity but walks past. This discrepancy is not discussed. The paper should clarify: (1) the trial termination protocol for OmniVLA (when does a non-stopping trial end — time limit, distance limit, operator intervention?), (2) whether OmniVLA's native inference stack was configured for the pose-goal modality in each environment, and (3) whether the goal coordinate frame was correctly aligned between the two systems. Without this, the reader cannot determine whether OmniVLA's failures reflect a real
  3. The claim that frozen MLLM visual features are 'sufficient' for robot navigation (Section 1, Section 8) is overgeneralized relative to the evidence. All four deployment environments and the SCAND training set share similar visual characteristics: outdoor/indoor structured spaces with paved surfaces, industrial settings, and parking areas. The TokenWalker OOD validation set (Appendix A) is described as covering 'outdoor paved access roads, sidewalks, and industrial yards' — visually similar to SCAND. The paper does not test fundamentally different visual domains (e.g., dense vegetation, dynamic crowded spaces, extreme lighting). The conclusion should scope the sufficiency claim to the tested visual regime, or the paper should include at least one deployment in a visually distinct environment.
minor comments (8)
  1. Section 3.2, Eq. 1: The bin index formula b(v) uses min(K-1, floor(...)) but does not specify behavior for v < -B (values below the range). The text mentions B=15m but the largest training goal arc-length is 14m (Appendix K), so this edge case may not arise in practice — still, a clamp note would help.
  2. Table 1: The 'Combined' column is the average of the three modalities but this is not stated in the caption. Clarify whether it is a weighted or simple average.
  3. Appendix G, Table 8: 'Thor' and 'Orin' are listed as embedded platforms but not further identified (model, manufacturer). Add brief hardware specifications.
  4. Section 4.4: The deployed policy is described as 'step-4 (final-phase, C2) checkpoint of the four-phase warm-start chain (Appendix P).' The four-phase chain construction (Appendix P, Q) is complex and the rationale for each phase is not immediately clear from the main text. A one-sentence summary of why the chain is structured this way would help the main-text reader.
  5. Appendix E.3: The crop-mode ablation reports that 'top' crop wins on OOD but 'stretch' wins in-distribution, and the paper deploys with 'stretch.' The rationale for this choice is not explicitly stated. Clarify why stretch was selected as the deployment default despite top winning OOD.
  6. Figure 1: The language prompt input is shown as 'grayed, not yet implemented.' This is honest but the figure caption should note that language conditioning is future work, not just visually grayed, to avoid confusion.
  7. Table 16 (Appendix N): The ego (image goal) row for GemNav shows SR as '1/4/0' but the Delta column shows '–'. If the model stopped once, a displacement value should be available. Clarify.
  8. The paper uses 'Gemma 4' throughout but this model is cited as [10] with a 2026 date. If this is a pre-release or non-public model, this should be noted for reproducibility.

Circularity Check

0 steps flagged

No circularity found. The derivation chain is self-contained and empirically grounded.

full rationale

The paper's central claims are empirical, not derivational. The soft-decoded auxiliary loss (Eqs. 3-4) is a standard weighted-expectation decoding over independently defined bin centers; it is not defined in terms of the quantity it predicts. The discrete token vocabulary (64 bins on [-15, 15]m) is specified independently of any output. The claim that frozen MLLM features suffice for navigation is tested by real-world deployment, not derived from a self-cited premise. The shared-token design for goal coordinates and outputs is a design choice whose benefit is measured empirically (Table 4: zero-shot step-2 ADE 0.43-0.46 vs specialized step-3 ADE 0.34-0.35), not asserted by definition. The only self-citation is [25] (Wildcat SLAM) for pose estimation, which is a localization tool, not a load-bearing theoretical premise. The TokenWalker dataset is author-collected but genuinely held out from training. No step in the chain reduces to its inputs by construction. The skeptic's concerns about OmniVLA baseline fairness are experimental-design issues, not circularity. This is a straightforward empirical paper with no circular reasoning detected.

Axiom & Free-Parameter Ledger

8 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical entities or postulated objects. The value-bin tokenization, soft-decoded loss, and special tokens (<goal_reached>, <goal_unreachable>, <bot_traj>, <eot_traj>, <goal_coord>) are methodological constructs, not invented entities with independent physical reality. All free parameters are standard hyperparameters for LoRA adaptation and tokenization, not fitted constants of nature.

free parameters (8)
  • K (value bins) = 64
    Number of discrete bins for waypoint coordinate quantization. Chosen so half-bin error (0.234m) is inside the 1m deployment success criterion.
  • B (bin range) = 15m
    Range of waypoint coordinates [-B, B]. Chosen to give 7% margin above largest training goal arc-length of 14m.
  • lambda_aux = 0.1
    Weight of soft-decoded auxiliary regression loss. Not tuned per modality; fixed throughout.
  • LoRA rank r = 32
    Low-rank adaptation dimension. Ablated against r=64 (Table 6) which was worse.
  • LoRA alpha = 16
    LoRA scaling factor. Fixed at 16, giving effective scaling alpha/r = 0.5.
  • N (waypoints) = 8
    Number of predicted waypoints per trajectory. Not ablated.
  • Learning rate = 1e-4
    AdamW learning rate with 500-step linear warmup.
  • Steps per phase = 40000
    Training budget per phase in the four-phase warm-start chain.
axioms (4)
  • domain assumption Pretrained MLLM visual features (from Gemma 4) are sufficiently general for robot navigation in structured environments
    The entire approach depends on the frozen vision tower providing adequate representations. This is tested empirically but not proven; the four deployment environments share visual characteristics with training data.
  • domain assumption Short-to-medium horizon waypoint navigation (8 waypoints, ~15m range) is a sufficient action space for the tested deployment scenarios
    The policy outputs 2D (x,y) waypoints in body frame, cannot express in-place rotation, and filters goals behind the robot (Section 7, Appendix K). This limits applicability to scenarios where these constraints are acceptable.
  • domain assumption Body-frame pose from LiDAR-inertial SLAM is accurate enough for goal specification and evaluation
    The pose-goal modality and all displacement metrics depend on SLAM-derived poses being accurate to sub-meter precision in all four environments.
  • domain assumption SCAND dataset trajectories are representative of deployable navigation behavior
    The policy learns entirely from SCAND teleoperation data; its deployment success depends on this data covering sufficient navigation patterns for the test environments.

pith-pipeline@v1.1.0-glm · 24562 in / 3367 out tokens · 257745 ms · 2026-07-09T23:36:28.551524+00:00 · methodology

0 comments
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 ↗

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 · 27 canonical work pages · 2 internal anchors

  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

    work page 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

    work page 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

    work page 2024

  4. [4]

    Embodied navigation foundation model.arXiv preprint arXiv:2509.12129, 2025

    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

    work page 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

    work page 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

    work page 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

    work page 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

    work page 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

    work page 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

    work page 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

    work page 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

    work page

  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

    work page 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

    work page 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

    work page 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

    work page

  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

    work page 2025

  18. [18]

    CAST: Counterfactual Labels Improve Instruction Following in Vision-Language-Action Models

    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

    work page internal anchor Pith review 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

    work page 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

    work page 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

    work page 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

    work page 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

    work page 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

    work page 2024

  25. [25]

    Wildcat: Online Continuous-Time 3D Lidar-Inertial SLAM

    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

    work page internal anchor Pith review 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

    work page 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) ...

    work page 2020