Pith. sign in

REVIEW 3 major objections 7 minor 36 references

VoLN: Vision-Only Long-Horizon Navigation---Paradigm, Benchmark, and Method

T0 review · 3 major / 7 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read This paper proposes replacing language-supplied route plans with in-scene visual cues for long-horizon navigation, instantiates a 7,210-episode aerial benchmark, and shows a strong reference baseline still succeeds only 7.4% of the time on

desk verdict A genuinely useful new benchmark interface for vision-only long-horizon aerial navigation, but the central claim that in-scene beacons are load-bearing is asserted rather than demonstrated. read the letter →

arxiv 2607.21400 v1 pith:XYHXPYDI submitted 2026-07-23 cs.RO cs.AI

Jiabin Lou , Haopeng Wang , Yuanshuai Wang , Xinyu Liu , Xuxin Lv , Yuxin Guo , Lei Huang , Rongye Shi
show 1 more author
Wenjun Wu
This is my paper
classification cs.ROcs.AI
keywords vision-onlynavigationlong-horizonaerialbenchmarksemanticbeaconsvisual-semanticalignmentclosed-loopplanningimage-goal
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

Vision-and-Language Navigation benchmarks typically hand the agent route structure—orientation, distance, layout—through instruction text, so scores conflate visual navigation with using externally supplied plans. The paper argues for a complementary interface, Vision-Only Long-Horizon Navigation (VoLN), where goal views specify the destination and route-relevant information exists only as locally observable in-scene cues such as semantic beacons. It instantiates VoLN-UAV, a 7,210-episode aerial benchmark with continuous 3D motion, large viewpoint changes, and both active beacons and passive distractors, and provides VoLN-MLLM as a reference baseline. VoLN-MLLM aligns a self-supervised visual backbone into a semantic embedding space and uses a frozen language-model planner to emit short-horizon waypoints in closed loop; it outperforms adapted instruction-following baselines but reaches only 7.4%, 4.5%, and 1.8% success on Test-Unseen Easy, Normal, and Hard episodes. If the benchmark is valid, it establishes a new evaluation axis—vision-only long-horizon aerial navigation—whose low success rates define a concrete open challenge.

What carries the argument

The load-bearing objects are the visual goal set V (the last three RGB frames of each reference trajectory, replacing language as the destination specifier), the active and passive semantic beacons (active beacons placed at decision points, passive beacons serving as fixed distractors, together forcing online detection, interpretation, and context-dependent selection), and the two-stage VoLN-MLLM pipeline. That pipeline consists of a lightweight adapter that distills a self-supervised visual backbone's features into a frozen vision-language embedding space, enabling top-k retrieval from a fixed semantic bank, followed by a frozen large-language-model planner (adapted with low-rank updates) t

What would settle it

Run a human pilot or an oracle policy that is given ground-truth identities and locations of active beacons on the Test-Unseen split, and compare its success rate with that of a strong goal-image-matching agent that ignores beacons; if the two are close, beacon selection is not the operative difficulty and the benchmark reduces to image-goal matching.

Watch

Extended reading notes

Core claim

The paper's central claim is that current VLN benchmarks measure a mixture of visual navigation and route structure provided by the instruction, and that a cleaner interface can isolate the navigation ability itself. In VoLN, the agent receives only egocentric RGB observations, proprioception, and a visual goal set (the final three frames of the reference trajectory); no GPS, global maps, or language instructions are available. VoLN-UAV operationalizes this as a 7,210-episode benchmark that embeds the cue-discrimination problem in the scene: active beacons at decision points provide task-relevant guidance, while roughly 150 passive beacons per environment act as visually similar distractors.

Load-bearing premise

The benchmark's claim to measure beacon-guided navigation rests on the unchecked premise that active beacons are visually distinguishable from passive look-alikes in egocentric RGB views and that success actually depends on selecting the right beacons, rather than on matching the goal images alone.

Editorial extensions

If this is right

  • Adopting the VoLN interface would let benchmark scores separate visual navigation skill from the spatial priors that language instructions leak, enabling cleaner comparisons across navigation methods.
  • The 7.4%, 4.5%, and 1.8% success rates on Test-Unseen quantify headroom and point directly to the bottlenecks the paper names: long-horizon evidence integration, cross-view goal matching, and closed-loop stability.
  • The active/passive beacon design creates a concrete testbed for cue selection under distractor clutter; an agent that solves VoLN-UAV must demonstrate it can ignore look-alike passive beacons.
  • The ablations show that the visual-semantic alignment stage is the largest single contributor in the reference method: removing it drops success from 5.7% to 2.3% on Test-Unseen, indicating that representation compatibility with the semantic space matters more than planner adaptation or encoder choice in this setup.
  • Because the interface consists only of egocentric RGB, goal views, and proprioception, it can be instantiated on physical platforms, as the paper demonstrates with a scaled indoor testbed rollout.

Reading between the lines

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

  • If the benchmark's construct validity holds, the same interface could transfer to other GPS-denied, language-free embodied settings—ground robots, underwater vehicles—where route-relevant information must be read from the environment rather than from a plan.
  • The low absolute success rates suggest the bottleneck may be memory and evidence accumulation rather than raw perception; a natural next test is adding an explicit spatial or topological memory to a VoLN-style agent and measuring the gain.
  • A stronger test of the benchmark's motivation would be an oracle that receives ground-truth active-beacon identities: if oracle success stays close to a goal-image-matching agent's success, the beacons are not yet load-bearing and the benchmark is closer to image-goal navigation than the authors intend.
  • The fixed semantic bank, retrieved by top-k cosine similarity, limits the vocabulary to descriptors chosen offline; an agent that builds task-specific cue vocabularies online might generalize better to unseen environments.
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

3 major / 7 minor

Summary. The paper proposes Vision-Only Long-Horizon Navigation (VoLN), a paradigm in which an agent receives only goal views and proprioception, with route-relevant information available only through locally observable in-scene cues such as semantic beacons. It introduces VoLN-UAV, a 7,210-episode aerial navigation benchmark built in Unreal Engine/AirSim with active and passive beacons, scene-source-based splits, and three difficulty strata. The paper also presents VoLN-MLLM, a two-stage method that aligns DINO visual features with a CLIP semantic space and uses a frozen LLM with LoRA to predict short-horizon waypoints and stop decisions. Experiments compare VoLN-MLLM against adapted visual-goal baselines and report the highest point estimates on all metrics, with Test-Unseen success rates of 7.4%, 4.5%, and 1.8% on Easy, Normal, and Hard episodes.

Significance. If the central claims hold, VoLN-UAV would be a useful complementary benchmark: it removes language-supplied route structure and tests long-horizon visual navigation in continuous 3D, an underexplored regime. The scene-source split is appropriate, the task interface is clearly defined, and the baseline method is a reasonable initial reference given the very low absolute success rates. The paper is appropriately modest in calling VoLN-MLLM an initial reference. However, the benchmark's core motivation—that the task measures 'context-dependent beacon selection' and that route-relevant information is locally observable—is not empirically substantiated. The paper also lacks basic statistical rigor in its headline comparisons. Both issues are addressable within the scope of a revision.

major comments (3)
  1. [§4.2 and §3] The benchmark's central construct-validity claim is unsupported. The paper asserts that active beacons are 'sparsely placed at decision points' and passive beacons have 'similar visual forms,' but provides no evidence that active beacons are visually distinguishable in egocentric RGB, that agents can detect them at decision points, or that using them is necessary for success. There are no beacon-detection metrics, no oracle-with-beacon-conditioned baseline, no experiment that removes or replaces active beacons with passive ones, and no human performance. Figure 6 is qualitative and only shows one trajectory passing through active-beacon locations; this could happen coincidentally for any trajectory following the reference route. Without these controls, the low success rates cannot be attributed to failure in 'detect, interpret, and select' cues rather than to image-goal matching under vi
  2. [Table 1 and §6.2] The headline comparisons are single-seed point estimates with no variance, confidence intervals, or significance testing. The Test-Unseen Hard subset contains roughly 12% of 1,081 episodes (~130 episodes); VoLN-MLLM's 1.8% SR versus LAG-VG's 0.4% thus corresponds to roughly 2 versus 0.5 successful episodes. The claimed 'relative advantage persists' is not established at this scale. Please report multiple seeds with standard deviations, bootstrap confidence intervals, or a significance test, and show raw episode counts for success on each difficulty level. This applies to the ablations in Table 2 as well, which also report only point estimates.
  3. [§4.2 and §6.1] The difficulty thresholds L1=300 m and L2=450 m and the success region radius ε=4 m are presented without justification or sensitivity analysis. The paper's Easy/Normal/Hard stratification and all success metrics depend on these choices. I ask for a sensitivity analysis (e.g., SR as a function of ε, or difficulty composition under alternative thresholds) and a statement of how these values relate to environment scale and task length. This is not the main blocker, but it is load-bearing for the benchmark's claimed difficulty taxonomy.
minor comments (7)
  1. [General] The project page is mentioned in the abstract but the paper does not state whether benchmark data, environment assets, evaluation scripts, and baseline code will be released under a persistent license. Please include a clear availability statement.
  2. [§4.2] 'Decision point' is used without a formal definition. Specify how decision points are identified, how many active beacons are visible per view, and how 'context-dependent' relevance is operationalized (i.e., whether an active beacon for one episode can be a passive distractor in another).
  3. [Eq. (3)] The choice of the final three observations as the visual goal set is arbitrary. Justify this choice and consider ablating the number of goal views, since the task difficulty may depend strongly on the amount of viewpoint variation covered by the goal set.
  4. [§6.1] The binary cross-entropy term in Eq. (6) is not defined inline. A short definition of BCE would improve readability.
  5. [Table 1] The multi-row header with nested columns is hard to read. Consider separating the table by split or using a longtable format with clearer grouping.
  6. [Abstract and text] There are typographical inconsistencies such as 'VoLN-UA V' with an extra space; unify the dataset name across the paper.
  7. [§6.3] The physical testbed demonstration is described as a single 'representative rollout' with no metrics. While this is clearly preliminary, adding quantitative descriptors (e.g., success/failure counts, path length, time) would strengthen it.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the benchmark construction and baseline evaluation are self-contained; the beacon-validity concern is a construct-validity threat, not a circular derivation.

full rationale

The paper's central deliverables are a new benchmark (VoLN-UAV), a reference baseline (VoLN-MLLM), and comparative evaluations. None of these reduce to a fitted parameter being renamed as a prediction. The loss functions in Eqs. 4-6 supervise the baseline against demonstrations taken from the same benchmark, but that is standard supervised training, not circular reasoning. The stop threshold τ is selected on Validation-Seen and then applied to Test-Unseen, which is a legitimate use of a development set. The claim that VoLN-MLLM outperforms adapted baselines is a point-estimate comparison on held-out environments, and the baselines share the same interface and training targets, so the comparison is not forced by construction. Self-citations appear only in related-work and environment-source contexts (e.g., TALKER, Air-M, AERIS) and are not load-bearing for the paradigm, benchmark, or method. The main substantive weakness is that the benchmark's central construct—context-dependent beacon selection—is not directly validated: the paper asserts active beacons are 'sparsely placed at decision points as task-relevant cues' and passive beacons provide 'semantic clutter,' but it does not demonstrate that success requires beacon selection or that beacons are visually distinguishable. This is an empirical validity threat, not a case where a prediction is equivalent to its inputs by construction. Therefore, the circularity score is minimal.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The benchmark's design introduces several hand-chosen constructs (beacons, goal-view interface, difficulty cuts) and the method adds tunable thresholds/hyperparameters. The paper does not provide external evidence for the most load-bearing of these (beacon detectability/usage), and several numerical values are unreported (τ, λ_stop, k).

free parameters (5)
  • stop threshold τ = not reported (selected on Validation-Seen)
    Chosen in §5/§6 to maximize Validation-Seen success rate; directly controls SR on Test-Unseen and is not reported.
  • success radius ε = 4 m
    Hand-set in §6.1; larger ε inflates SR and smaller ε deflates it; no sensitivity analysis.
  • difficulty thresholds L1, L2 = L1=300 m, L2=450 m
    Hand-chosen in §4.2 to stratify episodes into Easy/Normal/Hard; changes the difficulty mix across the benchmark.
  • waypoint horizon H and step budget T = H=8, T=128
    Architecture and evaluation design choices in §6.1 that trade off replan frequency and episode length.
  • loss weight λ_stop = not reported
    Appears in Eq. (6) but its value is not given.
assumptions (4)
  • domain assumption AirSim/Unreal Engine simulation is a sufficiently valid instantiation of open 3D navigation for the conclusions drawn.
    All benchmark results and ablations are generated in simulation (§4.1); a real-world demonstration is a single indoor rollout (§6.3).
  • domain assumption DINOv3 features can be aligned to the CLIP embedding space and this alignment improves navigation-relevant semantic grounding.
    Phase I (§5) relies on this transferable-alignment premise; supported by prior work (Talking to DINO) but not independently validated here beyond the ablation.
  • ad hoc to paper Active beacons are visually distinguishable from passive beacons and provide decision-point information that the agent can detect in egocentric RGB.
    Beacon augmentation (§4.2) assumes this property without rendering/visibility statistics or beacon-detection metrics; no experiment isolates beacon usage.
  • ad hoc to paper The final three observations of the reference trajectory (Eq. 3) specify the destination across the viewpoint variation expected during approach.
    Goal-view set V(ξ) in Eq. (3) is the only destination interface; the paper does not analyze viewpoint coverage or goal-region ambiguity.
invented entities (2)
  • Active and passive semantic beacons
    purpose: Provide route-relevant in-scene cues (active) and structured distractors (passive) to instantiate the VoLN task (§4.2).
    Their semantics are defined by the authors; there is no external data supporting their detectability from egocentric RGB or their role in navigation success.
  • Fixed semantic bank C
    purpose: Defines the structured semantic space into which visual features are aligned and from which tokens are retrieved for the planner (§5).
    The bank's category vocabulary is not specified, and retrieval quality is not evaluated; it is a free construct of the method.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VoLN: Vision-Only Long-Horizon Navigation---Paradigm, Benchmark, and Method." pith.science (2026). https://pith.science/paper/XYHXPYDI

@misc{pith2026260721400,
  author       = {Pith},
  title        = {Pith review of: VoLN: Vision-Only Long-Horizon Navigation---Paradigm, Benchmark, and Method},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XYHXPYDI}},
  note         = {Machine review of arXiv:2607.21400}
}
read the original abstract

Vision-and-Language Navigation (VLN) enables embodied agents to follow natural-language instructions. However, route-level instructions commonly encode spatial priors, such as orientation, distance, and layout, that are not explicitly available from onboard sensing at deployment in open, GPS-denied environments. Benchmark performance under such interfaces therefore jointly reflects visual navigation ability and the use of route structure explicitly supplied by the task description. As a complementary formulation, we propose Vision-Only Long-Horizon Navigation (VoLN), which shifts route-relevant information from externally supplied instructions and global guidance to locally observable in-scene cues. In VoLN, goal views specify the destination, while route-relevant information is available only through locally observable in-scene cues that the agent must detect, interpret, and select online. We instantiate VoLN for aerial navigation through VoLN-UAV, a 7,210-episode benchmark that combines long-horizon goal-directed flight, continuous 3D motion, large viewpoint changes, and context-dependent beacon selection. We further provide VoLN-MLLM as an initial reference baseline. It aligns self-supervised visual features with a structured semantic space and predicts short-horizon waypoint segments from observation history, goal views, retrieved visual--semantic tokens, and proprioception. On the five-environment Test-Unseen split, it obtains success rates of 7.4%, 4.5%, and 1.8% on Easy, Normal, and Hard episodes, respectively. These results provide an initial evaluation of VoLN and reveal substantial remaining challenges in long-horizon evidence integration, cross-view goal matching, and closed-loop stability. Project page: https://admire-ljb.github.io/VoLN-UAV/

Figures

Figures reproduced from arXiv: 2607.21400 by the authors.

Figure 1
Figure 1. The instruction-based setting exposes route-level [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The VoLN interaction. The policy maps goal views [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Representative VoLN-UAV environments. its construction stages and resulting data organization. Trajectory collection. Panels (a) and (e) relate reference￾route collection to the resulting path-length distribution. For each episode, we sample a scene from E and select a reference route from the corresponding route pool. The pool contains predefined routes from existing datasets and custom routes collected by trained … view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Overview of VoLN-UAV. Panels (a)–(d) show trajectory collection, beacon augmentation, annotation generation, and [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: VoLN-MLLM overview. Phase I learns visual– [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Successful and failed rollouts. Top: the trajectory passes the active-beacon locations and enters the goal region. Bottom: the trajectory diverges near the passive billboards and terminates outside the goal region [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Physical testbed and a representative VoLN rollout. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 5 canonical work pages

  1. [1]

    Proceedings of the AAAI Conference on Artificial Intelligence , year =

    Zhou, Gengze and Hong, Yicong and Wu, Qi , title =. Proceedings of the AAAI Conference on Artificial Intelligence , year =. doi:10.1609/aaai.v38i7.28597 , url =

    doi:10.1609/aaai.v38i7.28597

  2. [2]

    , title =

    Chen, Jiaqi and Lin, Bingqian and Xu, Ran and Chai, Zhenhua and Liang, Xiaodan and Wong, Kwan-Yee K. , title =. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =. doi:10.18653/v1/2024.acl-long.529 , url =

    work page doi:10.18653/v1/2024.acl-long.529 2024

  3. [3]

    2025 , eprint =

    Qi, Zhangyang and Zhang, Zhixiong and Yu, Yizhou and Wang, Jiaqi and Zhao, Hengshuang , title =. 2025 , eprint =

    2025

  4. [4]

    2025 , eprint =

    Lin, Sihao and Li, Zerui and Zhao, Xunyi and Zhou, Gengze and Wang, Liuyi and Wei, Rong and Tang, Rui and Li, Juncheng and Wang, Hanqing and Pang, Jiangmiao and van den Hengel, Anton and Liu, Jiajun and Wu, Qi , title =. 2025 , eprint =

    2025

  5. [5]

    International Conference on Learning Representations (ICLR) , year =

    Zhang, Jiazhao and Li, Anqi and Qi, Yunpeng and Li, Minghan and Liu, Jiahang and Wang, Shaoan and Liu, Haoran and Zhou, Gengze and Wu, Yuze and Li, Xingxing and Fan, Yuxin and Li, Wenjun and Chen, Zhibo and Gao, Fei and Wu, Qi and Zhang, Zhizheng and Wang, He , title =. International Conference on Learning Representations (ICLR) , year =

  6. [6]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    Liu, Shubo and Zhang, Hongsheng and Qi, Yuankai and Wang, Peng and Zhang, Yanning and Wu, Qi , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

  7. [7]

    International Conference on Learning Representations (ICLR) , year =

    Wang, Xiangyu and Yang, Donglin and Wang, Ziqin and Kwan, Hohin and Chen, Jinyu and Wu, Wenjun and Li, Hongsheng and Liao, Yue and Liu, Si , title =. International Conference on Learning Representations (ICLR) , year =

  8. [8]

    International Conference on Learning Representations (ICLR) , year =

    Gao, Yunpeng and Li, Chenhui and You, Zhongrui and Liu, Junli and Li, Zhen and Chen, Pengan and Chen, Qizhi and Tang, Zhonghan and Wang, Liansheng and Yang, Penghui and Tang, Yiwen and Tang, Yuhang and Liang, Shuai and Zhu, Songyi and Xiong, Ziqin and Su, Yifei and Ye, Xinyi and Li, Jianan and Ding, Yan and Wang, Dong and Wang, Zhigang and Zhao, Bin and L...

Show all 36 references
  1. [9]

    Drones , year =

    Lou, Jiabin and Ding, Rong and Wu, Wenjun , title =. Drones , year =. doi:10.3390/drones8060221 , url =

    doi

  2. [10]

    Champion-Level Drone Racing Using Deep Reinforcement Learning , journal =

    Kaufmann, Elia and Bauersfeld, Leonard and Loquercio, Antonio and M. Champion-Level Drone Racing Using Deep Reinforcement Learning , journal =. 2023 , volume =. doi:10.1038/s41586-023-06419-4 , url =

    2023 doi

  3. [11]

    Machine Learning , year =

    Krishnan, Srivatsan and Boroujerdian, Behzad and Fu, William and Faust, Aleksandra and Reddi, Vijay Janapa , title =. Machine Learning , year =. doi:10.1007/s10994-021-06006-6 , url =

    doi

  4. [12]

    2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , year =

    Lou, Jiabin and Wu, Wenjun and Liao, Shuhao and Shi, Rongye , title =. 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , year =. doi:10.1109/IROS55552.2023.10341405 , url =

    2023

  5. [13]

    2025 European Conference on Mobile Robots (ECMR) , year =

    Saxena, Pranav and Raghuvanshi, Nishant and Goveas, Neena , title =. 2025 European Conference on Mobile Robots (ECMR) , year =. doi:10.1109/ECMR65884.2025.11163198 , url =

    2025

  6. [14]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

    Zhang, Weichen and Gao, Chen and Yu, Shiquan and Peng, Ruiying and Zhao, Baining and Zhang, Qian and Cui, Jinqiang and Chen, Xinlei and Li, Yong , title =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =....

    2025 doi

  7. [15]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , year =

    Cai, Hengxing and Dong, Jinhan and Tan, Jingjun and Deng, Jingcheng and Li, Sihang and Gao, Zhifeng and Wang, Haidong and Su, Zicheng and Sumalee, Agachai and Zhong, Renxin , title =. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , year...

    2025 doi

  8. [16]

    2026 , eprint =

    Lou, Jiabin and Wang, Haopeng and Liu, Xinyu and Zhang, Yu and Shi, Rongye and Wu, Wenjun , title =. 2026 , eprint =. doi:10.48550/arXiv.2606.30151 , url =

    2026 doi arXiv

  9. [17]

    IEEE Robotics and Automation Letters , year =

    Lou, Jiabin and Shi, Rongye and Lin, Yuxin and Wang, Qunbo and Wu, Wenjun , title =. IEEE Robotics and Automation Letters , year =. doi:10.1109/LRA.2024.3511434 , url =

    2024

  10. [18]

    IEEE Transactions on Automation Science and Engineering , year =

    Lou, Jiabin and Shi, Rongye and Wang, Haopeng and Yu, Ming-Ming and Wang, Yuanshuai and Wang, Qunbo and Wu, Wenjun , title =. IEEE Transactions on Automation Science and Engineering , year =. doi:10.1109/TASE.2026.3689345 , url =

    2026

  11. [19]

    International Conference on Learning Representations (ICLR) , year =

    Bono, Guillaume and Antsfeld, Leonid and Chidlovskii, Boris and Weinzaepfel, Philippe and Wolf, Christian , title =. International Conference on Learning Representations (ICLR) , year =

  12. [20]

    2022 , eprint =

    Krantz, Jacob and Lee, Stefan and Malik, Jitendra and Batra, Dhruv and Chaplot, Devendra Singh , title =. 2022 , eprint =. doi:10.48550/arXiv.2211.15876 , url =

    2022 doi arXiv

  13. [21]

    2024 , eprint =

    Pelluri, Nikhilanj , title =. 2024 , eprint =

    2024

  14. [22]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year =

    Lei, Xiaohan and Wang, Min and Zhou, Wengang and Li, Houqiang , title =. IEEE Transactions on Pattern Analysis and Machine Intelligence , year =. doi:10.1109/TPAMI.2025.3538496 , url =

    2025

  15. [26]

    2024 IEEE International Conference on Robotics and Automation (ICRA) , year =

    Cai, Wenzhe and Huang, Siyuan and Cheng, Guangran and Long, Yuxing and Gao, Peng and Sun, Changyin and Dong, Hao , title =. 2024 IEEE International Conference on Robotics and Automation (ICRA) , year =. doi:10.1109/ICRA57147.2024.10610499 , url =

    2024

  16. [28]

    Proceedings of The 8th Conference on Robot Learning , year =

    Zhang, Mike and Qu, Kaixian and Patil, Vaishakh and Cadena, Cesar and Hutter, Marco , title =. Proceedings of The 8th Conference on Robot Learning , year =

  17. [29]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Zhang, Sixian and Yu, Xinyao and Song, Xinhang and Wang, Xiaohan and Jiang, Shuqiang , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =. doi:10.1109/CVPR52733.2024.01553 , url =

    2024

  18. [31]

    2025 IEEE International Conference on Robotics and Automation (ICRA) , year =

    Anwar, Abrar and Welsh, John and Biswas, Joydeep and Pouya, Soha and Chang, Yan , title =. 2025 IEEE International Conference on Robotics and Automation (ICRA) , year =. doi:10.1109/ICRA55743.2025.11127706 , url =

    2025

  19. [32]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Vision-and-Language Navigation: Interpreting Visually-Grounded Navigation Instructions in Real Environments , author =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =. doi:10.1109/CVPR.2018.00387 , url =

    2018

  20. [33]

    2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , year =

    Nie, Dujun and Guo, Xianda and Duan, Yiqun and Zhang, Ruijun and Chen, Long , title =. 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , year =. doi:10.1109/IROS60139.2025.11246684 , url =

    2025

  21. [34]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , year =

    Shah, Hardik and Xing, Jiaxu and Messikommer, Nico and Sun, Boyang and Pollefeys, Marc and Scaramuzza, Davide , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , year =. doi:10.1109/CVPRW67362.2025.00519 , url =

    2025

  22. [35]

    2025 , eprint =

    Huang, Yanjia and Wu, Mingyang and Li, Renjie and Tu, Zhengzhong , title =. 2025 , eprint =

    2025

  23. [36]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    Barsellotti, Luca and Bianchi, Lorenzo and Messina, Nicola and Carrara, Fabio and Cornia, Marcella and Baraldi, Lorenzo and Falchi, Fabrizio and Cucchiara, Rita , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =. doi:10.1109/ICCV...

    2025

  24. [37]

    and Zhang, Hao and Gonzalez, Joseph E

    Zheng, Lianmin and Chiang, Wei-Lin and Sheng, Ying and Zhuang, Siyuan and Wu, Zhanghao and Zhuang, Yonghao and Lin, Zi and Li, Zhuohan and Li, Dacheng and Xing, Eric P. and Zhang, Hao and Gonzalez, Joseph E. and Stoica, Ion , title =. Advances in Neural Information Processing ...

    2020 doi

  25. [38]

    and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , title =

    Hu, Edward J. and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , title =. International Conference on Learning Representations (ICLR) , year =

  26. [39]

    2018 , eprint=

    On Evaluation of Embodied Navigation Agents , author=. 2018 , eprint=

    2018

  27. [40]

    Computer Vision -- ECCV 2020 , year =

    Krantz, Jacob and Wijmans, Erik and Majumdar, Arjun and Batra, Dhruv and Lee, Stefan , title =. Computer Vision -- ECCV 2020 , year =. doi:10.1007/978-3-030-58604-1_7 , url =

    2020 doi

  28. [41]

    NeurIPS Visually Grounded Interaction and Language (ViGIL) Workshop , year =

    Ilharco, Gabriel and Jain, Vihan and Ku, Alexander and Ie, Eugene and Baldridge, Jason , title =. NeurIPS Visually Grounded Interaction and Language (ViGIL) Workshop , year =

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.