Anticipate Before Acting: Future-State-Conditioned Vision-Language Navigation

Liheng Ma; Lingfeng Zhang; Tongtong Cao; Yingxue Zhang; Zhanguang Zhang

REVIEW 3 major objections 4 minor 13 references

A training-only future-state token makes a vision-language navigator anticipate, improving long-horizon success by up to 3–4 points.

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 · deepseek-v4-flash

2026-08-01 16:15 UTC pith:6D4OTVQG

load-bearing objection Plausible VLN future-conditioning paper whose headline gains rest on single runs and whose 'Δ-step' target actually reduces to the segment endpoint; worth refereeing if the authors fix both issues. the 3 major comments →

arxiv 2607.18042 v2 pith:6D4OTVQG submitted 2026-07-20 cs.CV cs.AI

Anticipate Before Acting: Future-State-Conditioned Vision-Language Navigation

Lingfeng Zhang , Zhanguang Zhang , Liheng Ma , Tongtong Cao , Yingxue Zhang This is my paper

classification cs.CV cs.AI

keywords vision-language navigationfuture-state predictionauxiliary supervisioncausal vision-language modelbehavior cloninglong-horizon navigationlatent alignmentunseen environment generalization

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.

This paper asks whether a navigation policy that predicts the next action can be improved by also forcing part of its internal state to anticipate a future observation. It first shows, in a diagnostic that feeds the expert's future image as a privileged input, that future visual evidence dramatically helps action choice. It then introduces a deployable training scheme: a learnable future-query token is aligned, only during training, to a frozen visual embedding of the observation Δ steps ahead, and the resulting state conditions action generation through causal attention. On an unseen-environment benchmark, the method improves success-rate metrics by roughly one to three percentage points overall, with larger gains on trajectories longer than 10 meters. The significance is that predictive supervision can help an end-to-end navigator without changing inference-time inputs or adding a future-image generator.

Core claim

The paper's central claim is that standard behavior cloning, which supervises only the next action, leaves the policy hidden state unconstrained with respect to future visual outcomes; adding a future-state alignment objective during training makes part of the state anticipate a future visual latent, and this improves navigation, especially over long horizons. The implementable mechanism inserts two learnable prefix tokens, a future query followed by an action query, before action tokens in a causal vision-language model. During training, the hidden state at the future query is pushed by a projected L1 loss with stop-gradient toward the mean-pooled visual embedding, computed by the same froz

What carries the argument

The central object is the dual-query causal prefix: two learnable tokens, a future query qF and an action query qA, inserted between the multimodal context and the autoregressive action tokens. During training, qF's contextualized hidden state is aligned, through a projected L1 loss with stop-gradient, to the mean-pooled visual embedding (from the same frozen encoder) of the expert trajectory's observation Δ steps ahead; qA then attends to qF, creating an explicit causal path from future-state prediction to action decoding. This arrangement lets the policy use a predictive internal state without seeing the future at inference, at the cost of only two prefix tokens.

Load-bearing premise

The gains rest on the assumption that aligning the future-query state to the expert trajectory's future-frame embedding (usually the segment's final frame) teaches a genuinely predictive representation; if the target is essentially just the segment endpoint, or if the expert future diverges from what the agent actually encounters, the auxiliary loss may not transfer to deployment.

What would settle it

Train the same model using a future target that is guaranteed to be a true Δ-step-ahead observation — for example, sampled from a held-out corridor beyond the segment or from the same scene at a fixed offset that never coincides with the segment end — and compare against the current implementation; if the gains vanish, the improvement comes from endpoint prediction rather than anticipation. Alternatively, run the whole training procedure with multiple seeds and check whether the 0.98–1.47 and 2.93–3.85 percentage-point gains lie outside the seed noise floor.

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

If this is right

Future-state supervision can be added to existing causal VLM navigators with negligible overhead: two prefix tokens, one auxiliary loss, and no change to inference-time inputs or decoding.
Gains concentrated on episodes longer than 10 meters suggest that predictive state supervision matters most where error accumulation is severe, making horizon-stratified evaluation important for such methods.
The dual-query separation implies that future-supervised states should not be directly reused as action-conditioning states; a dedicated action query preserves the predictive signal for decoding.
Because the target branch is frozen and training-only, the mechanism composes with larger data regimes and could combine with other future-latent distillation schemes.
The diagnostic result — privileged access to a future image lifts success rate dramatically — indicates substantial headroom in future-aware navigation, encouraging further work on predictive training signals.

Where Pith is reading between the lines

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

Since Δ equals the maximum segment length (32 steps) and the target falls back to the final frame when a trajectory ends, the auxiliary loss may often reduce to 'predict the segment endpoint' rather than true constant-offset anticipation; a test with genuinely out-of-segment future targets would determine which effect drives the gains.
The reported gains of one to four percentage points come from single runs, so run-to-run variance could account for part of the effect; repeated-seed evaluation would sharpen the claim.
The paper's own limitation — expert-future targets may differ from the agent's on-policy future under compounding errors — suggests a natural extension: adaptive or uncertainty-weighted targets, or using the policy's own rollouts to form targets, could amplify the benefit.
Because the method operates at the token level with a frozen visual encoder, it could transfer to other instruction-following embodied tasks, such as manipulation, where long-horizon error accumulation is similarly limiting.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Plausible VLN future-conditioning paper whose headline gains rest on single runs and whose 'Δ-step' target actually reduces to the segment endpoint; worth refereeing if the authors fix both issues.

read the letter

Two things to know. First, the paper's central empirical claim—that future-state supervision at a fixed horizon improves navigation—is not as clean as stated: with training segments capped at 32 steps and Δ=32, the target is almost always the final frame of the segment, so the loss supervises endpoint prediction, not Δ-step anticipation. Second, all headline gains are single runs with no error bars or significance tests, and the aggregate improvements are 1–1.5pp SR—within plausible run-to-run noise for VLN training. The long-horizon gains (3–4pp) are more interesting but share the same weakness.

What is genuinely good: the privileged-input diagnostic is a nice sanity check—feeding the expert future frame at train/test lifts SR from 30% to 72%—and it motivates training-only supervision cleanly. The method itself is simple and cheap: two learned prefix tokens, a frozen SigLIP target with stop-gradient, and an ℓ1 loss in a low-dim projection. That matches established practice from the manipulation literature (VLA-JEPA, FastWAM, etc.), which the paper cites honestly. The long-horizon stratification is the right way to look for an effect, and the authors are upfront about limitations: expert-future mismatch, fixed Δ, single backbone. Credit where due: the writing is clear and the method is internally consistent.

Where it wobbles. The Δ=32/segment-length issue is not a footnote: it changes what the method actually does. Instead of 'predict the observation 32 low-level steps ahead,' the loss is 'predict the endpoint of this training segment.' That may still be a useful predictive signal, and it may even explain the long-horizon benefit better than fixed-horizon anticipation, but it is not what the paper claims, and it affects how generalizable the result is. The single-run problem is real: at 1pp aggregate gains, you need variance or repeated seeds to believe the effect exists at all. The long-horizon numbers are bigger, but without error bars they are still one roll of the dice. Minor issue: the Discussion says the action query helps Oracle SR, but Table 4 does not report OSR; that claim is unsupported in the table. No code or artifacts are released, which would help with the variance concern.

Who it's for: people working on VLN or future-conditioned VLAs will want to read it, mainly as a transfer attempt and a cautionary example of supervision targets not matching the stated horizon. It deserves a serious referee, but only with explicit request for repeated-seed experiments and a corrected description of what the target actually is.

Referee Report

3 major / 4 minor

Summary. The paper proposes FSC-VLN, a training-only auxiliary future-state supervision method for end-to-end vision-language navigation. A StreamVLN-style causal Qwen2-7B policy is augmented with two learnable prefix tokens, a future query qF and an action query qA. During training, the hidden state at qF is aligned, via an L1 loss in a lower-dimensional projection space, to the mean-pooled, stop-gradient SigLIP embedding of an expert-trajectory observation Δ=32 low-level steps ahead; the target branch is discarded at inference. The paper first reports a privileged-input diagnostic in which the expert future image is given as an additional input at train and test time, improving R2R val-unseen SR from 29.96% to 71.77%. It then reports deployable gains over the baseline: +1.47 SR / +3.27 OSR / +1.11 SPL under R2R+RxR training and +0.98 / +1.04 / +0.85 under R2R+RxR+EnvDrop+ScaleVLN, with larger long-horizon gains (e.g., +3.85 SR under the first regime). An ablation removes qA and shows a small drop in SR/SPL.

Significance. If the reported effect is real, the contribution is a useful and remarkably lightweight training-time intervention: a small auxiliary loss that forces part of the policy hidden state to anticipate a frozen future visual embedding, with negligible inference overhead and no future input at deployment. The method is clearly motivated by the privileged-input diagnostic, and the architecture is coherent: the causal path qF→qA→actions is explicit, the target is frozen/stop-gradient, and the paper honestly discloses the expert-future/on-policy mismatch and the single-backbone scope. The stratified long-horizon analysis is also a sensible way to test the mechanism. However, the central empirical claim rests on single-run numbers with no error bars, and the future-target construction is ambiguous in a way that could change the interpretation of the loss. These issues are load-bearing for the paper's main claim that future-state anticipation—rather than a generic endpoint-prediction or summarization signal—improves long-horizon navigation.

major comments (3)

[Sec. 4, Implementation details; Eq. (18)] The future-target construction is underspecified in a way that determines what the loss actually supervises. The text says each training sample is a trajectory segment of up to 32 low-level control steps and that the target is at offset Δ=32, with the final frame used if t+Δ exceeds the trajectory length. If the target observation is looked up within the same training segment, then for every position t in the segment, t+Δ is at or beyond the segment boundary, so the loss reduces to predicting the segment endpoint/final frame rather than a fixed 32-step-ahead observation. If the target is instead drawn from the full expert trajectory beyond the sampled segment, then the fixed-horizon interpretation is correct. The current wording is ambiguous. Please specify precisely where ot+Δ is taken from (segment vs. full trajectory), give the sampling procedure, and, if the endpoint interpretation i
[Tables 2–3 and Sec. 4, Main Results] All reported gains are from single runs, with no error bars, confidence intervals, or significance tests. The aggregate gains are small (0.98–1.47 pp SR), and VLN training is known to exhibit run-to-run variance on this scale. The long-horizon stratified gains (2.93–3.85 pp SR) are more substantial but are also single-run values. Please report mean and standard deviation over at least three seeds, or provide bootstrap confidence intervals or paired significance tests over episodes, for both the aggregate and the stratified results. Without this, the central empirical claim is not statistically supported. The same concern applies to the ablation in Table 4, where the differences are even smaller.
[Sec. 3.6, Eq. (18)] The projection g is not defined as fixed or learned, and no training details are given for it. If g is a learned projection, the stop-gradient placed on z* does not prevent g from exploiting the objective: because g is applied to both the predicted and target quantities, the L1 loss can in principle be minimized by collapsing g's output (e.g., driving it toward zero), even without the future-query representation being predictive. The paper should specify whether g is a fixed random projection, a frozen projection, or a learned projection, and, if learned, explain why the action-prediction loss prevents this collapse. If a learned projection with a stop-gradient after the target projection is intended, state that explicitly.

minor comments (4)

[Abstract / Introduction] Several formatting artifacts appear in the PDF text, e.g., "Aprivileged-input diagnosticshows" and "FastW AM". Please proofread.
[Abstract vs. Sec. 4] The abstract says "StreamVLN-style baseline" while the experiments section says "We use StreamVLN as our baseline." Clarify whether the baseline is the official StreamVLN model or a re-implemented StreamVLN-style model, and, if re-implemented, report how the reproduction was validated.
[Eq. (13)] The superscript in h^{act}_{t,j−1} appears malformed in the text; it should denote the hidden state of the preceding action token.
[Sec. 3.6] The projection dimension d′ is not specified in the experiments. Please report it, along with the initialization of qF and qA (the paper says N(μ, σ²) but does not give μ or σ).

Circularity Check

0 steps flagged

No significant circularity: the auxiliary future-state loss is a genuine stop-gradient target and the central claim is evaluated on an external benchmark.

full rationale

After walking the derivation chain, I find no circular step. The only construction that resembles a prediction target is Eq. (18), L_future = ||g(hat z_{t+Delta}) - g(z*_{t+Delta})||_1, with z*_{t+Delta} = sg[mean_pool V(o_{t+Delta})] (Eqs. 15-16). This target is a frozen, deterministic embedding of an expert future observation; it is not a function of the policy's own outputs, the baseline's metrics, or any fitted parameter. The policy's navigation quality is evaluated on R2R val-unseen by SR/OSR/SPL, which are external to the training loss. The privileged-input diagnostic (Table 1) is explicitly labeled as a motivation, not a claimed deployable result, and it does not define the auxiliary target. The use of Delta=32 with segments of up to 32 steps is an empirical concern (the target can reduce to the segment endpoint) but it does not make the derivation circular: the loss still aligns a policy state to a stop-gradient visual embedding from a future expert frame. The paper's own limitation statement that the expert future 'may differ from the agent's on-policy future' is an acknowledged assumption, not a circular reduction. The only overlapping self-citation, H-WM [Huang et al., 2026], appears in related work as motivation and is not load-bearing for the paper's results. Therefore score 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

FSC-VLN is an empirical method: its central claims rest on hyperparameters (λ_F=0.1, Δ=32, projection dimension d'), a set of domain assumptions about frozen-encoder latent spaces and expert trajectories as targets, and two new learnable tokens. No deep mathematical axioms are introduced; the ledger reflects choices and premises needed for the reported gains to be meaningful.

free parameters (4)

future-loss weight λ_F = 0.1
Chosen by hand; no sensitivity analysis reported. Eq. (21), Sec. 4.
future offset Δ = 32 steps
Chosen by hand; the paper lists fixed-Δ as a limitation in Sec. 5. Sec. 4.
projection dimension d' = not reported
g: R^d → R^d' in Eq. (18); the dimension is never given, though it determines the alignment loss geometry.
query initialization and placement = N(µ,σ²) with µ,σ unspecified; one future + one action query
Learned query embeddings are an architectural choice validated only by the single dual-query ablation (Table 4).

axioms (4)

domain assumption Frozen SigLIP + MLP defines a stationary, meaningful target space for future-state alignment (Eq. 15–16).
The usefulness of L_future presumes mean-pooled frozen embeddings of future frames carry actionable visual information; §3.5.
domain assumption The causal path qF→qA→action-token gradients shapes qF usefully (Eq. 8–9, §3.6).
The paper assumes action-prediction gradients flowing through the future query make it both predictive and action-useful; not proven separately from the ablation.
domain assumption Expert-trajectory future frames are valid training targets for an on-policy policy.
Acknowledged in Sec. 5 Limitations: expert future may differ from on-policy future under compounding errors.
standard math Standard Transformer causal attention and pretrained Qwen2-7B weights function as assumed (Eq. 7).
Background architecture assumptions; no formal verification.

invented entities (2)

future-query token qF no independent evidence
purpose: Learned prefix token whose contextualized hidden state is aligned to a frozen embedding of an expert future frame during training (Eq. 8, 17–18).
Its value is evidenced only by in-paper experiments; no external falsifiable handle.
action-query token qA no independent evidence
purpose: Learned prefix token that conditions first action decoding on the future-supervised state (Eq. 9, 12).
Supported only by the in-paper ablation (Table 4).

pith-pipeline@v1.3.0-alltime-deepseek · 9255 in / 18092 out tokens · 179946 ms · 2026-08-01T16:15:19.507252+00:00 · methodology

0 comments

read the original abstract

End-to-end vision-language navigation (VLN) with causal vision-language models maps instructions and egocentric observations directly to actions, but standard behavior cloning supervises only the next action and does not explicitly encourage the policy state to be predictive of future visual outcomes, limiting long-horizon decision making. A privileged-input diagnostic shows that access to an expert-trajectory future image can substantially improve navigation, indicating that future observations contain rich, actionable cues, though such inputs are unavailable at deployment. Motivated by this signal, we propose Future-State-Conditioned VLN (FSC-VLN), a deployable model that augments a causal policy with a future-query token and uses training-only future-state supervision to distill information from future observations into the policy state. Concretely, during training we align the future-query representation to a frozen visual embedding $\Delta$ steps ahead, while inference requires only past and current observations. This design preserves the baseline inference pattern and adds only two learned prefix tokens, implying minimal overhead. On R2R val-unseen, FSC-VLN improves SR/OSR/SPL over a StreamVLN-style baseline under two training-data regimes, with larger gains on long-horizon episodes; ablations further support the dual-query design that separates future and action queries.

Figures

Figures reproduced from arXiv: 2607.18042 by Liheng Ma, Lingfeng Zhang, Tongtong Cao, Yingxue Zhang, Zhanguang Zhang.

discussion (0)

Reference graph

Works this paper leans on

13 extracted references · 8 linked inside Pith

[4]

Causal world modeling for robot control.arXiv preprint arXiv:2601.21998,

Lin Li, Qihang Zhang, Yiming Luo, Shuai Yang, Ruilin Wang, Fei Han, Mingrui Yu, Zelin Gao, Nan Xue, Xing Zhu, et al. Causal world modeling for robot control.arXiv preprint arXiv:2601.21998,

Pith/arXiv arXiv
[6]

Learning to navigate unseen environments: Back translation with environmental dropout

Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, V olume 1 (Long and Short Papers), pages 2610–2621,

2019
[9]

Qwen2 technical report.arXiv preprint arXiv:2407.10671,

An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report.arXiv preprint arXiv:2407.10671,

Pith/arXiv arXiv
[10]

World action models are zero-shot policies.arXiv preprint arXiv:2602.15922,

Seonghyeon Ye, Yunhao Ge, Kaiyuan Zheng, Shenyuan Gao, Sihyun Yu, George Kurian, Suneel Indupuru, You Liang Tan, Chuning Zhu, Jiannan Xiang, et al. World action models are zero-shot policies.arXiv preprint arXiv:2602.15922,

Pith/arXiv arXiv
[11]

Fast-wam: Do world action models need test-time future imagination?arXiv preprint arXiv:2603.16666,

Tianyuan Yuan, Zibin Dong, Yicheng Liu, and Hang Zhao. Fast-wam: Do world action models need test-time future imagination?arXiv preprint arXiv:2603.16666,

Pith/arXiv arXiv
[12]

Janusvln: Decoupling semantics and spatiality with dual implicit memory for vision-language navigation.arXiv preprint arXiv:2509.22548,

Shuang Zeng, Dekang Qi, Xinyuan Chang, Feng Xiong, Shichao Xie, Xiaolong Wu, Shiyi Liang, Mu Xu, Xing Wei, and Ning Guo. Janusvln: Decoupling semantics and spatiality with dual implicit memory for vision-language navigation.arXiv preprint arXiv:2509.22548,

arXiv
[13]

Uni-navid: A video-based vision-language-action model for unifying embodied navigation tasks.arXiv preprint arXiv:2412.06224, 2024a

10 Jiazhao Zhang, Kunyu Wang, Shaoan Wang, Minghan Li, Haoran Liu, Songlin Wei, Zhongyuan Wang, Zhizheng Zhang, and He Wang. Uni-navid: A video-based vision-language-action model for unifying embodied navigation tasks.arXiv preprint arXiv:2412.06224, 2024a. Jiazhao Zhang, Kunyu Wang, Rongtao Xu, Gengze Zhou, Yicong Hong, Xiaomeng Fang, Qi Wu, Zhizheng Zha...

Pith/arXiv arXiv
[2018]

Navila: Legged robot vision-language-action model for navigation.arXiv preprint arXiv:2412.04453,

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.arXiv preprint arXiv:2412.04453,

Pith/arXiv arXiv
[2020]

Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding

Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4392–4412,

2020
[2023]

Streamvln: Streaming vision-and-language navigation via slowfast context modeling.arXiv preprint arXiv:2507.05240,

Meng Wei, Chenyang Wan, Xiqian Yu, Tai Wang, Yuqiang Yang, Xiaohan Mao, Chenming Zhu, Wenzhe Cai, Hanqing Wang, Yilun Chen, et al. Streamvln: Streaming vision-and-language navigation via slowfast context modeling.arXiv preprint arXiv:2507.05240,

Pith/arXiv arXiv
[2024]

H-wm: Robotic task and motion planning guided by hierarchical world model.arXiv preprint arXiv:2602.11291,

Jinbang Huang, Wenyuan Chen, Zhiyuan Li, Oscar Pang, Xiao Hu, Lingfeng Zhang, Yuanzhao Hu, Zhanguang Zhang, Mark Coates, Tongtong Cao, et al. H-wm: Robotic task and motion planning guided by hierarchical world model.arXiv preprint arXiv:2602.11291,

arXiv
[2025]

From foundation to application: Improving vla models in practice.arXiv preprint arXiv:2607.06403,

Wei Wu, Fangjing Wang, Fan Lu, He Sun, Shi Liu, Yunnan Wang, Yibin Yan, Yong Wang, Shuailei Ma, Xinyang Wang, et al. From foundation to application: Improving vla models in practice.arXiv preprint arXiv:2607.06403,

Pith/arXiv arXiv
[2026]

Vla-jepa: Enhancing vision-language-action model with latent world model

Jingwen Sun, Wenyao Zhang, Zekun Qi, Shaojie Ren, Zezhi Liu, Hanxin Zhu, Guangzhong Sun, Xin Jin, and Zhibo Chen. Vla-jepa: Enhancing vision-language-action model with latent world model. arXiv preprint arXiv:2602.10098,

arXiv

[1] [4]

Causal world modeling for robot control.arXiv preprint arXiv:2601.21998,

Lin Li, Qihang Zhang, Yiming Luo, Shuai Yang, Ruilin Wang, Fei Han, Mingrui Yu, Zelin Gao, Nan Xue, Xing Zhu, et al. Causal world modeling for robot control.arXiv preprint arXiv:2601.21998,

Pith/arXiv arXiv

[2] [6]

Learning to navigate unseen environments: Back translation with environmental dropout

Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, V olume 1 (Long and Short Papers), pages 2610–2621,

2019

[3] [9]

Qwen2 technical report.arXiv preprint arXiv:2407.10671,

An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report.arXiv preprint arXiv:2407.10671,

Pith/arXiv arXiv

[4] [10]

World action models are zero-shot policies.arXiv preprint arXiv:2602.15922,

Seonghyeon Ye, Yunhao Ge, Kaiyuan Zheng, Shenyuan Gao, Sihyun Yu, George Kurian, Suneel Indupuru, You Liang Tan, Chuning Zhu, Jiannan Xiang, et al. World action models are zero-shot policies.arXiv preprint arXiv:2602.15922,

Pith/arXiv arXiv

[5] [11]

Fast-wam: Do world action models need test-time future imagination?arXiv preprint arXiv:2603.16666,

Tianyuan Yuan, Zibin Dong, Yicheng Liu, and Hang Zhao. Fast-wam: Do world action models need test-time future imagination?arXiv preprint arXiv:2603.16666,

Pith/arXiv arXiv

[6] [12]

Janusvln: Decoupling semantics and spatiality with dual implicit memory for vision-language navigation.arXiv preprint arXiv:2509.22548,

Shuang Zeng, Dekang Qi, Xinyuan Chang, Feng Xiong, Shichao Xie, Xiaolong Wu, Shiyi Liang, Mu Xu, Xing Wei, and Ning Guo. Janusvln: Decoupling semantics and spatiality with dual implicit memory for vision-language navigation.arXiv preprint arXiv:2509.22548,

arXiv

[7] [13]

Uni-navid: A video-based vision-language-action model for unifying embodied navigation tasks.arXiv preprint arXiv:2412.06224, 2024a

10 Jiazhao Zhang, Kunyu Wang, Shaoan Wang, Minghan Li, Haoran Liu, Songlin Wei, Zhongyuan Wang, Zhizheng Zhang, and He Wang. Uni-navid: A video-based vision-language-action model for unifying embodied navigation tasks.arXiv preprint arXiv:2412.06224, 2024a. Jiazhao Zhang, Kunyu Wang, Rongtao Xu, Gengze Zhou, Yicong Hong, Xiaomeng Fang, Qi Wu, Zhizheng Zha...

Pith/arXiv arXiv

[8] [2018]

Navila: Legged robot vision-language-action model for navigation.arXiv preprint arXiv:2412.04453,

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.arXiv preprint arXiv:2412.04453,

Pith/arXiv arXiv

[9] [2020]

Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding

Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4392–4412,

2020

[10] [2023]

Streamvln: Streaming vision-and-language navigation via slowfast context modeling.arXiv preprint arXiv:2507.05240,

Meng Wei, Chenyang Wan, Xiqian Yu, Tai Wang, Yuqiang Yang, Xiaohan Mao, Chenming Zhu, Wenzhe Cai, Hanqing Wang, Yilun Chen, et al. Streamvln: Streaming vision-and-language navigation via slowfast context modeling.arXiv preprint arXiv:2507.05240,

Pith/arXiv arXiv

[11] [2024]

H-wm: Robotic task and motion planning guided by hierarchical world model.arXiv preprint arXiv:2602.11291,

Jinbang Huang, Wenyuan Chen, Zhiyuan Li, Oscar Pang, Xiao Hu, Lingfeng Zhang, Yuanzhao Hu, Zhanguang Zhang, Mark Coates, Tongtong Cao, et al. H-wm: Robotic task and motion planning guided by hierarchical world model.arXiv preprint arXiv:2602.11291,

arXiv

[12] [2025]

From foundation to application: Improving vla models in practice.arXiv preprint arXiv:2607.06403,

Wei Wu, Fangjing Wang, Fan Lu, He Sun, Shi Liu, Yunnan Wang, Yibin Yan, Yong Wang, Shuailei Ma, Xinyang Wang, et al. From foundation to application: Improving vla models in practice.arXiv preprint arXiv:2607.06403,

Pith/arXiv arXiv

[13] [2026]

Vla-jepa: Enhancing vision-language-action model with latent world model

Jingwen Sun, Wenyao Zhang, Zekun Qi, Shaojie Ren, Zezhi Liu, Hanxin Zhu, Guangzhong Sun, Xin Jin, and Zhibo Chen. Vla-jepa: Enhancing vision-language-action model with latent world model. arXiv preprint arXiv:2602.10098,

arXiv