Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Can Pretrained Vision-Language Embeddings Alone Guide Robot Navigation?

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that frozen vision-language embeddings alone, with a behavior-cloned policy, reach language-specified targets 74% of the time in a simulated arena, versus 100% for a state-aware expert in about a third of the steps.

desk verdict A clean, narrow empirical baseline whose headline claim overreaches because the directional words in the instruction leak privileged spatial information, and the paper never ablates that confound. read the letter →

arxiv 2506.14507 v1 pith:ADYJK2SR submitted 2025-06-17 cs.RO

Nitesh Subedi , Adam Haroon , Shreyan Ganguly , Samuel T.K. Tetteh , Prajwal Koirala , Cody Fleming , Soumik Sarkar This is my paper
classification cs.RO
keywords Vision-LanguageModelsBehaviorCloningLanguage-GuidedNavigationFrozenEmbeddingsFoundationRobotLearning
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

The paper asks a deliberately stripped-down question: if you take a pretrained vision-language embedding and add nothing else—no fine-tuning, no map, no memory, no spatial module—can a learned policy use it to follow a natural-language navigation command? To answer, the authors train a privileged expert with full state access, record its demonstrations, convert each camera image and instruction into a frozen joint embedding, and fit a behavior-cloning policy to the expert's actions. In a 100-episode evaluation inside a simple simulated arena with five colored spheres, the policy reaches the language-specified target 74% of the time, versus 100% for the expert, and successful runs take about 3.2 times more steps. The sympathetic reading is that frozen embeddings genuinely ground color language in vision for easy target selection, but they do not supply the spatial reasoning or long-horizon planning needed to navigate efficiently. A reader should care because the result sets a bare-bones baseline for when a robotics system can get away with a frozen embedding and when it must add specialized machinery.

What carries the argument

The mechanism that carries the argument is the frozen joint embedding. At each timestep, the camera image and the language instruction are passed separately through SigLIP's pretrained encoders; the two output vectors are L2-normalized, summed, and re-normalized into a single 1152-dimensional vector. A feedforward policy network then maps that vector to two wheel-velocity commands by minimizing squared error against the expert's recorded actions. Before training, the paper also deploys a spatial-sensitivity probe that measures cosine distance between joint embeddings of images with the same object in the same or different 3 × 3 grid cells, and uses that probe to choose SigLIP over CLIP and ViLT; navigation success then tracks the probe's ranking.

What would settle it

Retrain the same pipeline with each instruction's directional phrase replaced by a random or deliberately wrong one while keeping the color word intact; if success stays near 74%, the visual embedding itself is doing the spatial grounding, and if success collapses toward the reported color-only level, the hand-computed directional template carries the result.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that a behavior-cloned policy fed only a frozen joint image-text embedding can perform basic language-guided navigation: in a 3 m by 3 m simulated arena with five colored spheres, the SigLIP-based policy reaches the language-specified target in 74 of 100 test episodes, while a privileged PPO expert with full state access succeeds in all 100. Successful runs by the learned policy take on average 369.4 timesteps, about 3.2 times the expert's 114.0. The same training procedure with CLIP reaches 62% and with ViLT 40%, a ranking the paper traces to the models' measured spatial sensitivity in the joint embedding space rather than to behavior-cloning loss. The authors also report that color-only instructions perform much worse than instructions carrying relative directional phrases, and they read the overall gap as evidence that frozen embeddings supply semantic grounding but not the spatial memory and planning needed for efficient paths.

Load-bearing premise

The instructions handed to the robot contain directional phrases such as 'to your left' that are computed from the target's position relative to the robot's starting heading, so the spatial information credited to the visual embeddings may actually be entering through the prompt instead.

Editorial extensions

If this is right

  • For unobstructed 'go to the [color] object' tasks, a frozen vision-language embedding plus a behavior-cloned policy can serve as a functioning drop-in, with no VLM fine-tuning.
  • A cheap embedding-space probe—separating same-position from different-position pairs of images—can predict which VLM will navigate best before any robot training is done.
  • Behavior-cloning loss should not be used to compare policies for embodied tasks; in these experiments the model with the highest cloning loss navigated best.
  • Systems that need short paths, reliable exploration, or disambiguation of visually similar targets should expect to add spatial memory, mapping, or task-specific adaptation on top of frozen embeddings.

Reading between the lines

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

  • Beyond the paper: feeding the policy a short window of past joint embeddings—rather than a single frame—is a natural next test, since the observed circling and timeout failures suggest the policy lacks temporal continuity more than visual grounding.
  • Beyond the paper: the same spatial-sensitivity probe could serve as a general model-selection filter for other embodied tasks, but the paper only demonstrates the correlation in one colored-sphere arena, so that transfer remains a hypothesis.
  • Beyond the paper: because the directional phrases in the prompts are computed from privileged geometry, the 74% figure should be read as 'frozen embeddings plus hand-engineered directional language'; transferring the same prompt template to a different VLM or environment is needed before calling it a property of embeddings alone.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper asks whether frozen pretrained vision-language embeddings, used as the sole representation for a behavior-cloned navigation policy, can guide a robot to language-specified targets. The authors collect demonstrations in NVIDIA Isaac Sim from a privileged PPO expert with full state access, encode each RGB image and a template-generated language instruction with SigLIP into a joint L2-normalized embedding, and train a feedforward behavior-cloning policy on the expert actions. On 100 test episodes in a 3m x 3m arena with five colored spheres, the learned policy achieves 74% success versus 100% for the expert, using on average 3.2x more timesteps. The paper also compares ViLT, CLIP, and SigLIP, reporting success rates of 40%, 62%, and 74%, and interprets the efficiency gap as evidence that pretrained embeddings support semantic grounding but are insufficient for efficient spatial reasoning and long-horizon planning. Code is made available at https://github.com/oadamharoon/text2nav.

Significance. If the attribution is correct, the paper provides a useful empirical baseline: a minimalist student-teacher pipeline in which frozen VLM embeddings alone yield nontrivial language-guided navigation without fine-tuning, mapping modules, or specialized spatial components. The multi-model comparison, the privileged-expert reference point, and the public code release are concrete strengths, and the 74% success rate is a directly measured quantity. However, the central claim that the VLM's visual embeddings are what drive navigation is under-supported: the language instruction contains a directional cue computed from the target's ground-truth position relative to the robot's starting heading, and the paper itself reports in Section V-B that color-only prompts perform much worse. Without an ablation that removes or scrambles this directional information, the measured success may reflect the policy exploiting a privileged spatial label in the text channel rather than visual grounding by frozen embeddings.

major comments (4)
  1. [Section V-B and Section II-A] The directional prompt is a load-bearing confound. Instructions are generated with a spatial phrase ('to your left', 'to your right', 'straight ahead') computed from the target's position relative to the robot's starting heading (Section II-A), and the paper reports that color-only prompts perform 'significantly lower' (Section V-B). Because the directional token is constant for the whole episode, the behavior-cloned policy can learn a reactive strategy that mostly follows this static directional label and then switches to color once the target is in view. The experiment therefore does not yet show that vision-language embeddings alone provide the spatial grounding for navigation; it may instead show that privileged spatial cues inserted into the language channel help. The missing control is an ablation that removes, randomizes, or inverts the directional text while keeping the frozen VLM pipeline otherwise identical, reporting both success rate and timesteps. This ablation is necessary to support the paper's central attribution.
  2. [Section II-B] The VLM selection probe may measure text similarity rather than visual spatial encoding. In the spatial sensitivity test, the joint embeddings are computed from images paired with spatial language prompts, and images in different spatial cells plausibly receive different directional words. The observed lower cosine distance for 'same cell' pairs could therefore reflect the text encoder assigning similar vectors to identical spatial words, rather than the image encoder encoding object position. The authors should repeat the analysis with a fixed text prompt across all image pairs, or with image-only embeddings, to substantiate the claim that SigLIP has stronger inherent spatial sensitivity in its joint embedding space. This matters because the selection of SigLIP and the later 'spatial sensitivity ranking mirrors navigation performance' argument rest on this measurement.
  3. [Section III and Section IV] The evaluation lacks the minimal baselines needed to attribute the 74% success to the joint vision-language representation. No random-policy baseline, image-only (no text) policy, text-only (no image) policy, or color-only-prompt policy is reported. Given the small arena, a random or heading-following policy may achieve nonzero success simply by moving forward and colliding with a sphere, and the authors' own Section V-B note that color-only prompts perform much lower. Adding these controls would quantify how much of the 74% comes from the directional label, how much from visual object recognition, and how much from the joint embedding itself. Without them, the phrase 'embeddings alone guide navigation' is not established by the data.
  4. [Section IV and evaluation protocol] The statistical robustness of the main claim is unclear. The paper reports success rates for ViLT, CLIP, and SigLIP with 95% confidence intervals, but it does not state how many training runs or random seeds underlie the expert and student policies, and the main expert-versus-student comparison appears to be a single deployment of each policy on 100 episodes. The sensitivity of the results to the success threshold (0.1 m) and the episode horizon (1000 timesteps) is also not examined. Reporting the mean and variance over at least a few training seeds, and a small sweep over the success radius and horizon, would materially strengthen the claim that the 74% versus 100% gap and the 3.2x step ratio are stable findings.
minor comments (6)
  1. [Section II-B] The sentence 'across x different configurations' leaves x undefined; the number of configurations and images used in the spatial sensitivity test should be stated explicitly.
  2. [Algorithm 1 and Section II-C] The notation 'L2Norm' is used without a formal definition, and the joint embedding is described as the L2-normalized sum of two L2-normalized vectors; the paper should state the normalization convention once in the text or in the algorithm caption.
  3. [Section I and Section II-C] Reference [1] is cited for SigLIP, but reference [1] is the 'Getting ViT in shape' scaling-laws paper; the SigLIP architecture and its 1152-dimensional embedding should be cited to reference [23].
  4. [Section IV-A] The claim that higher embedding dimensionality 'translates directly to better navigation performance' is confounded by differences in pretraining data, architecture, and training objective across ViLT, CLIP, and SigLIP; the authors should soften this causal phrasing.
  5. [Section IV and Figure 6] The behavioral cloning loss is averaged over 100 epochs without stating the batch size, network architecture, or whether the reported loss is on a held-out validation split; adding these details would improve reproducibility.
  6. [Section II-A] The dataset stores reward signals based on distance to goal, but these rewards are not used by the behavior-cloning loss; the paper should clarify whether these stored rewards are simply unused artifacts of the data collection process.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the navigation results are measured empirically against a separately trained privileged expert, and no fitted parameter or self-citation chain forces the main 74% claim.

full rationale

The paper's central result is an empirical measurement: a behavior-cloned policy trained on frozen SigLIP embeddings achieves 74% success while a privileged expert reaches 100%. Nothing in the method defines the student's success rate in terms of the inputs; the expert demonstrations are collected independently, and the student policy is evaluated on 100 test episodes with randomized starts. The language prompts do embed privileged directional cues (e.g., 'to your left') computed from the target's position relative to the robot's heading, and Section V-B admits that color-only prompts performed significantly worse. This is a genuine confound that weakens the claim that VLM embeddings alone provide spatial guidance, but it is not circularity: the 74% number is not derived from the prompt construction, and no equation in the paper reduces a prediction to an input by definition. The VLM selection probe in Section II-B measures cosine distances between joint embeddings under same/different spatial cells; SigLIP is chosen because it shows the largest separation. The later success hierarchy (ViLT < CLIP < SigLIP) is a post-hoc correlation, not a quantity fitted from the probe. No self-citation is load-bearing, and no uniqueness theorem or ansatz is imported from the authors' prior work. The evaluation is self-contained against an external benchmark (the privileged expert), so per the review rules this is an honest non-finding with score 0. The prompt-sensitivity confound should be recorded as a correctness risk, not as circularity.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

No new physical or mathematical entities are introduced. The central claim rests on standard pretrained models, a trained MLP policy, and hand-designed prompts; the main free choices are the prompt template and the evaluation thresholds.

free parameters (3)
  • Spatial prompt template = "The target is the [color] ball which is to your [left/right/straight ahead]. Move toward the ball."
    Hand-designed language template injects relative position into the text embedding; Section V-B states color-only prompts performed significantly worse, making this template load-bearing for the 74% result.
  • Spatial grid resolution for VLM selection = 3x3 grid
    Used in Section II-B to rank ViLT, CLIP, and SigLIP by spatial sensitivity; this ranking motivates the SigLIP choice and is later used to explain the navigation performance hierarchy.
  • Success threshold = 0.1 m
    Task definition in Section III(c); success rate depends on this radius, although it is a standard distance threshold rather than an ad hoc fitted quantity.
assumptions (3)
  • domain assumption Isaac Sim dynamics and rendering are a faithful proxy for real robot navigation behavior.
    All experiments are in NVIDIA Isaac Sim (Section III); no real-world validation is provided, so the transferability of the 74% number rests on simulation fidelity.
  • ad hoc to paper The L2-normalized sum of image and text embeddings is a valid joint representation for policy learning.
    Algorithm 1 lines 17-19 define the joint embedding as L2Norm(VLM_img(I) + VLM_text(l)); this fusion is chosen without comparison to concatenation or other pooling methods.
  • domain assumption A 3x3 meter empty arena with five colored spheres is a meaningful test of language-guided navigation.
    Section III(a) describes no obstacles and fixed colored spheres; conclusions about 'basic language grounding' are scoped to this simplified setting, and the paper does not test clutter or varied layouts.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Can Pretrained Vision-Language Embeddings Alone Guide Robot Navigation?." pith.science (2026). https://pith.science/paper/ADYJK2SR

@misc{pith2026250614507,
  author       = {Pith},
  title        = {Pith review of: Can Pretrained Vision-Language Embeddings Alone Guide Robot Navigation?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ADYJK2SR}},
  note         = {Machine review of arXiv:2506.14507}
}
read the original abstract

Foundation models have revolutionized robotics by providing rich semantic representations without task-specific training. While many approaches integrate pretrained vision-language models (VLMs) with specialized navigation architectures, the fundamental question remains: can these pretrained embeddings alone successfully guide navigation without additional fine-tuning or specialized modules? We present a minimalist framework that decouples this question by training a behavior cloning policy directly on frozen vision-language embeddings from demonstrations collected by a privileged expert. Our approach achieves a 74% success rate in navigation to language-specified targets, compared to 100% for the state-aware expert, though requiring 3.2 times more steps on average. This performance gap reveals that pretrained embeddings effectively support basic language grounding but struggle with long-horizon planning and spatial reasoning. By providing this empirical baseline, we highlight both the capabilities and limitations of using foundation models as drop-in representations for embodied tasks, offering critical insights for robotics researchers facing practical design tradeoffs between system complexity and performance in resource-constrained scenarios. Our code is available at https://github.com/oadamharoon/text2nav

Figures

Figures reproduced from arXiv: 2506.14507 by the authors.

Figure 2
Figure 2. Expert policy training reward. Reward plot for the expert policy πβ, which quickly learns to success￾fully navigate to specified targets using privileged state information. searchers seeking to balance simplicity and performance in language-guided robot navigation systems. Our contributions include: • A minimalist framework for language-guided navi￾gation that uses frozen vision-language embeddings as the sole repre… view at source ↗
Figure 3
Figure 3. Simulation environment. Simulation environ￾ment showing the Jetbot with multiple colored spheres. During evaluation, the robot must navigate to a spe￾cific colored sphere based on language instructions that include both semantic color identification and relative spatial cues using only RGB camera input and pretrained vision-language embeddings. We optimize this loss using Adam with an initial learning rate of 1 × 10… view at source ↗
Figure 4
Figure 4. Reward comparison. Cumulative rewards com￾parison between the expert (πβ, blue) and VLM-based policy (π, orange). The learned policy shows greater variability and consistently lower rewards due to inef￾ficient navigation and occasional failures. instructions in visual observations and generates ap￾propriate actions to reach the target—all through the lens of a frozen vision-language model and a simple feedforward ne… view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Multi-model performance comparison. Success rates across three vision-language models using identical BC training procedures (n=100 episodes each). SigLIP demonstrates superior performance, achieving 74.0% success rate compared to CLIP’s 62.0% and ViLT’s 40.0%. impact …
Figure 6
Figure 6. Figure 6: Behavioral cloning loss across VLM architec￾tures. SigLIP achieves superior navigation performance despite higher BC loss than ViLT and CLIP, suggesting that BC loss does not predict downstream task perfor￾mance for embodied applications. A. Semantic Grounding vs. Spat…

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. A Comprehensive Survey and Systematic Real-World Evaluation of Embodied Vision-and-Language Navigation

    cs.RO 2026-07 accept novelty 5.5 of 10

    VLN methods show a large sim-to-real gap; a hierarchical system reaches 51% real-world success versus 22% for a monolithic RGB-only system across ten physical scenes.

Reference graph

Works this paper leans on

26 extracted references · 10 canonical work pages · cited by 1 Pith paper

  1. [1]

    Getting vit in shape: Scaling laws for compute-optimal model design,

    Ibrahim Alabdulmohsin, Xiaohua Zhai, Alexander Kolesnikov, and Lucas Beyer. Getting vit in shape: Scaling laws for compute-optimal model design,

    work page

  2. [2]

    Spatialvlm: Endow- ing vision-language models with spatial reasoning capabilities, 2024

    Boyuan Chen, Zhuo Xu, Sean Kirmani, Brian Ichter, Danny Driess, Pete Florence, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endow- ing vision-language models with spatial reasoning capabilities, 2024. URL https://arxiv.org/abs/2401. 12168

    work page 2024

  3. [3]

    Navila: Legged robot vision-language-action model for navigation, 2025

    An-Chieh Cheng, Yandong Ji, Zhaojing Yang, Zaitian Gongye, Xueyan Zou, Jan Kautz, Erdem Bıyık, Hongxu Yin, Sifei Liu, and Xiaolong Wang. Navila: Legged robot vision-language-action model for navigation, 2025. URL https://arxiv.org/abs/ 2412.04453

    arXiv 2025

  4. [4]

    Sukhatme

    Vishnu Sashank Dorbala, Gunnar Sigurdsson, Robinson Piramuthu, Jesse Thomason, and Gau- rav S. Sukhatme. Clip-nav: Using clip for zero- shot vision-and-language navigation, 2022. URL https://arxiv.org/abs/2211.16649

    arXiv 2022

  5. [5]

    Learning latent dynamics for planning from pixels, 2019

    Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels, 2019. URL https://arxiv.org/abs/1811. 04551

    work page 2019

  6. [6]

    Mastering diverse domains through world models, 2024

    Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models, 2024. URL https://arxiv. org/abs/2301.04104

    arXiv 2024

  7. [7]

    Visual language maps for robot navigation, 2023

    Chenguang Huang, Oier Mees, Andy Zeng, and Wolfram Burgard. Visual language maps for robot navigation, 2023. URL https://arxiv.org/abs/2210. 05714

    work page 2023

  8. [8]

    Bc-z: Zero-shot task general- ization with robotic imitation learning, 2022

    Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kap- pler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task general- ization with robotic imitation learning, 2022. URL https://arxiv.org/abs/2202.02005

    arXiv 2022

Show all 26 references
  1. [9]

    Vilt: Vision-and-language transformer without convolu- tion or region supervision, 2021

    Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolu- tion or region supervision, 2021. URL https://arxiv. org/abs/2102.03334

    2021 arXiv

  2. [10]

    Distilling realizable students from unrealizable teachers, 2025

    Yujin Kim, Nathaniel Chin, Arnav Vasudev, and Sanjiban Choudhury. Distilling realizable students from unrealizable teachers, 2025. URL https:// arxiv.org/abs/2505.09546

    2025 arXiv

  3. [11]

    Constrained behavior cloning for robotic learning, 2024

    Wensheng Liang, Jun Xie, Zhicheng Wang, Jianwei Tan, and Xiaoguang Ma. Constrained behavior cloning for robotic learning, 2024. URL https: //arxiv.org/abs/2408.10568

    2024 arXiv

  4. [12]

    ZSON: Zero-shot object-goal navigation using multimodal goal embeddings

    Arjun Majumdar, Gunjan Aggarwal, Bhavika Suresh Devnani, Judy Hoffman, and Dhruv Batra. ZSON: Zero-shot object-goal navigation using multimodal goal embeddings. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Process...

    2022

  5. [13]

    Orbit: A unified simulation framework for interactive robot learning environ- ments

    Mayank Mittal, Calvin Yu, Qinxi Yu, Jingzhou Liu, Nikita Rudin, David Hoeller, Jia Lin Yuan, Ritvik Singh, Yunrong Guo, Hammad Mazhar, Ajay Man- dlekar, Buck Babich, Gavriel State, Marco Hutter, and Animesh Garg. Orbit: A unified simulation framework for interactive robot lear...

    2023 doi

  6. [14]

    Language-conditioned offline rl for multi-robot navigation, 2024

    Steven Morad, Ajay Shankar, Jan Blumenkamp, and Amanda Prorok. Language-conditioned offline rl for multi-robot navigation, 2024. URL https: //arxiv.org/abs/2407.20164

    2024 arXiv

  7. [15]

    R3m: A uni- versal visual representation for robot manipulation,

    Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A uni- versal visual representation for robot manipulation,

  8. [16]

    Ta, Baoru Huang, Thieu V o, Ngan Le, and Anh Nguyen

    Nghia Nguyen, Minh Nhat Vu, Tung D. Ta, Baoru Huang, Thieu V o, Ngan Le, and Anh Nguyen. Robotic-clip: Fine-tuning clip on action data for robotic applications, 2024. URL https://arxiv.org/ abs/2409.17727

    2024 arXiv

  9. [17]

    Isaac sim

    NVIDIA. Isaac sim. https://developer.nvidia.com/ isaac-sim, 2022

    2022

  10. [18]

    Learning transferable visual models from natural language supervision, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. URL https://ar...

    2021 arXiv

  11. [19]

    Latent plans for task-agnostic offline reinforcement learn- ing, 2022

    Erick Rosete-Beas, Oier Mees, Gabriel Kalweit, Joschka Boedecker, and Wolfram Burgard. Latent plans for task-agnostic offline reinforcement learn- ing, 2022. URL https://arxiv.org/abs/2209.08959

    2022 arXiv

  12. [20]

    Proximal policy optimization algorithms, 2017

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms, 2017. URL https://arxiv. org/abs/1707.06347

    2017 arXiv

  13. [21]

    Vlm-social-nav: Socially aware robot navigation through scoring using vision-language models, 2024

    Daeun Song, Jing Liang, Amirreza Payandeh, Amir Hossain Raj, Xuesu Xiao, and Dinesh Manocha. Vlm-social-nav: Socially aware robot navigation through scoring using vision-language models, 2024. URL https://arxiv.org/abs/2404. 00210

    2024

  14. [22]

    Open- world object manipulation using pre-trained vision- language models, 2023

    Austin Stone, Ted Xiao, Yao Lu, Keerthana Gopalakrishnan, Kuang-Huei Lee, Quan Vuong, Paul Wohlhart, Sean Kirmani, Brianna Zitkovich, Fei Xia, Chelsea Finn, and Karol Hausman. Open- world object manipulation using pre-trained vision- language models, 2023. URL https://arxiv.or...

    2023 arXiv

  15. [23]

    Sigmoid loss for language image pre-training, 2023

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training, 2023. URL https://arxiv.org/abs/2303.15343

    2023 arXiv

  16. [24]

    RT-2: Vision- language-action models transfer web knowledge to robotic control

    Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, Quan Vuong, Vincent Van- houcke, Huong Tran, Radu Soricut, Anikait Singh, Jaspiar Singh, Pierre Sermanet, Pannag R San- keti, Grecia Salazar, Michael S R...

    2023

  17. [2022]

    URL https://arxiv.org/abs/2203.12601

    arXiv

  18. [2024]

    URL https://arxiv.org/abs/2305.13035

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.