Pith. sign in

REVIEW 4 major objections 4 minor 23 references

Discovering Diverse Planning Policies for Multimodal Embodied Agents with Quality-Diversity Optimization

T0 review · 4 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read This paper claims that long-horizon embodied agents fail mainly because they lock onto a single planning style, and that maintaining a behavior-indexed archive of diverse planning-policy templates—built by quality-diversity search and…

desk verdict A genuinely new QD-over-prompt-templates approach with online stall switching, but the experimental reporting leaves the headline gains unverified because the evaluation split is unstated and all results are point estimates. read the letter →

arxiv 2608.08523 v1 pith:JXKECGKH submitted 2026-08-09 cs.AI

Pengfei Xu , Yong Liu , Xiaoya Nan , Qiang Yang , Peilan Xu This is my paper · ORCID
classification cs.AI
keywords LLMagentspromptpolicyquality-diversityMAP-Elitesembodiedplanningmultimodalinteractionbehaviorallock-inswitching
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 identifies behavioral lock-in—an agent sticking to one planning style even when that style has stopped producing progress—as a main cause of long-horizon failure in multimodal embodied agents. It proposes a two-stage remedy: offline, evolve a diverse archive of planning-policy templates (prompt scaffolds with editable rule modules) using quality-diversity search, indexing each policy by interaction intensity and goal-directedness; online, execute one policy, detect stalls as sustained lack of milestone progress, roll back to the latest checkpoint, and switch to the archive policy whose behavior is farthest from the current one. On the ThreeDWorld transport benchmark under a 2000-step cap, the full system raises food success from 0.33 (rule-based RHP) or 0.42 (CoELA with GPT-4) to 0.51, and stuff success from 0.24/0.36 to 0.43, with lower token consumption than CoELA. A pilot on VLN-CE navigation shows the same recovery mechanism improving a frozen base policy from 35.2% to 37.3% success. The sympathetic reading is that robustness in embodied planning comes from preserving several useful ways of acting and knowing when to switch among them, not from making any single policy more elaborate.

What carries the argument

The central object is a behavior-indexed archive of planning-policy templates, built by MAP-Elites quality-diversity search. Each policy is a template $\tau = (r_{\text{sense}}, r_{\text{progress}}, r_{\text{act}}, r_{\text{recover}})$ whose editable rule modules dictate sensing, progress interpretation, action commitment, and recovery style. Policies are mapped into a two-dimensional behavior space via $\mathrm{BD}(\tau) = (b_{\text{int}}, b_{\text{goal}})$, where $b_{\text{int}}$ is the fraction of information-gathering actions and $b_{\text{goal}}$ measures how consistently the policy turns interaction into milestone-advancing progress. The archive is a $10\times 10$ grid over this space; each cell retains the elite policy for its niche. The online mechanism that carries the argument is Eq. (17): on stall, switch to the archive policy maximizing normalized descriptor distance from the current policy, among candidates not yet tried since the last checkpoint. This farthest-policy rule is what turns offline diversity into a recovery resource.

What would settle it

Run the full system with the behavior descriptors replaced by random values while keeping everything else fixed. If task success does not drop appreciably, the claim that behavior-aware switching (Eq. 17) is the causal mechanism for recovery is falsified. A softer test: shuffle the descriptor coordinates in the archive and check whether the chosen farthest policies still produce distinguishable rollout trajectories; if they do not, the descriptor space is not capturing what the paper says it captures.

Watch

Extended reading notes

Core claim

The paper argues that long-horizon failures in embodied agents often come from behavioral lock-in: a single planning style that keeps producing locally plausible but globally unproductive actions. To address this, it proposes a two-stage framework. Offline, a MAP-Elites-style quality-diversity search evolves planning-policy templates—prompt scaffolds with editable rule modules for sensing, progress interpretation, action commitment, and recovery—and organizes them into an archive indexed by two behavior descriptors: interaction intensity (the fraction of information-gathering actions) and goal-directedness (how consistently interactions become milestone progress). Each niche stores the highest-quality policy found, where quality mixes success rate and efficiency. Online, the agent executes one policy, maintains a checkpoint at each milestone, detects stalls as sustained non-advancement of the progress state (10 consecutive steps), and upon stall rolls back to the last checkpoint and switches to the archive policy whose behavior descriptor is farthest from the current one (normalized Euclidean distance in the 2D behavior space). On the TDW-MAT transport benchmark with a 2000-step cap, the full system with GPT-4 reaches 0.51 food and 0.43 stuff success, versus 0.33/0.24 for the rule-based RHP baseline and 0.42/0.36 for CoELA with GPT-4, using lower average token consumption (62.36k vs 69.42k). A pilot on VLN-CE navigation with a frozen CMA policy plus a separate recovery archive similarly improves success from 35.2% to 37.3%, suggesting the mechanism transfers beyond object transport.

Load-bearing premise

The load-bearing premise is that the two hand-designed behavior descriptors—interaction intensity and goal-directedness—capture the behavioral differences that actually determine whether a new policy can escape a stall; if the descriptor space is uninformative, the farthest-policy switching rule (Eq. 17) could pick a policy that is behaviorally no more useful than the stalled one.

Editorial extensions

If this is right

  • If diverse policy repertoires are what drive robustness, single-prompt optimization methods such as ReAct, Reflexion, and Tree-of-Thoughts should be expected to plateau on long-horizon embodied tasks where behavioral lock-in dominates.
  • The same archive can serve multiple tasks without retraining: the VLN-CE pilot shows a recovery archive built from navigation rollouts transfers to a different domain (waypoint following) and improves a frozen base policy.
  • Because the quality-diversity variant with LLaMA-2 uses fewer tokens than CoELA with GPT-4 (51.21k) while achieving competitive or better success, diversity-based recovery need not increase inference cost.
  • Ablations show the archive alone is much weaker than the full system (0.37/0.25 vs 0.51/0.43 without vs with online switching), so the stall-triggered switching and rollback mechanism are what convert offline diversity into online gains.

Reading between the lines

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

  • A natural extension is learned behavior descriptors: instead of hand-designing interaction intensity and goal-directedness, derive descriptors from trajectory embeddings or task-progress signals so the archive generalizes to settings where the right recovery behavior is not known in advance.
  • The distance-maximizing switch rule treats the archive as a discrete repertoire; a continuous interpolation between policy templates could allow finer-grained behavioral control between niches, potentially improving recovery when the farthest policy overshoots the needed behavior.
  • The stall detector is a fixed threshold on progress-state stasis; a learned or adaptive detector that anticipates stalls from trajectory statistics could switch earlier and save steps, provided it does not cause oscillation between policies.
  • The pilot result suggests a modular design: a frozen base policy plus a separate recovery archive can be added to an existing agent without retraining the base, making the mechanism a plug-in robustness layer for deployed embodied agents.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a Quality-Diversity (QD) framework for multimodal embodied planning. Offline, planning-policy templates are evolved with a MAP-Elites-style search and organized into an archive indexed by two hand-defined behavior descriptors (interaction intensity and goal-directedness, Eqs. 4 and 5). Online, the agent executes a single policy, detects stalls via a fixed threshold on progress-state stagnation (Eq. 15), rolls back to a checkpoint, and switches to the behaviorally farthest untried archive elite (Eq. 17). Experiments on TDW-MAT report improved food/stuff success over RHP and CoELA baselines, and a small VLN-CE pilot is used to argue that the recovery mechanism transfers across domains. The central empirical claims are that the full method raises Food success from 0.33 (RHP) and 0.42 (CoELA) to 0.51, and Stuff from 0.24/0.36 to 0.43, under a 2000-step budget.

Significance. If the empirical claims were well supported, the paper would make a useful contribution: it applies QD search to LLM prompt-level planning policies in a way that is clearly described, releases code, and includes an ablation separating online switching from archive selection. The online rollback-and-switch mechanism is a sensible response to behavioral lock-in, and the idea of treating behavioral diversity as a recoverable resource is worth exploring. However, the current evaluation does not yet establish the central claim. The absence of any statement that archive-construction rollouts are disjoint from evaluation episodes, the lack of error bars or significance tests, and the underspecified ablations all prevent verification of the reported gains. The VLN-CE pilot is too thinly described to support the cross-domain transfer claim. With additional experiments and clarification, the core idea may be salvageable, but the evidence as presented is not sufficient.

major comments (4)
  1. [Section 3.3.2, Algorithm 1; Section 4.1] The manuscript never states that the episodes used for offline rollout evaluation (Algorithm 1, line 14, which produces Q(τ) in Eq. (10) and BD(τ) in Eqs. (4)–(5)) are disjoint from the episodes scored in Tables 1, 2, and 4. If the archive is constructed on the same 'standard food and stuff splits' used for evaluation, then the QD pipeline has been effectively selected on the test set, whereas RHP and CoELA have not; the headline improvements (e.g., 0.33 to 0.51 on Food, 0.24 to 0.43 on Stuff) would then be an artifact of evaluation-set selection rather than evidence for diversity-driven robustness. Please specify the exact episode split between offline rollouts and online evaluation, and if no such split exists, re-run with a disjoint held-out split.
  2. [Section 4.1, Tables 1–4] All reported successes are point estimates with no seeds, standard deviations, or significance tests. For example, the VLN-CE SR increase from 35.2% to 37.3% (Table 3) and the Food gain from 0.42 (CoELA) to 0.51 (Ours) in Table 1 are within the range of plausible noise for a single run. Without multiple independent runs or episode-level bootstrap confidence intervals, the magnitude of the claimed improvement is unquantified, and statements such as 'the improvement is consistent across both task categories' (Section 4.2) are not supportable. Please report variance or confidence intervals for all main tables.
  3. [Section 4.3, Table 4] The ablation 'Ours (no OS)' is underspecified: when stall-based switching is disabled, it is unclear which policy is followed for the entire episode—the highest-quality archive elite, a random archive entry, or the initial seed template. The 'OS + Abl. ArchSel' condition is also ambiguous because it is described as relying on 'a simplified criterion (quality-only, or an equivalent random elite choice)'; these two choices can give very different baselines. Without a precise selection rule, the ablation cannot support the causal claims in Section 4.3 that online switching and behavior-aware selection each contribute to the observed gains.
  4. [Section 4.2, Table 3] The VLN-CE pilot is presented as evidence that the recovery mechanism transfers across domains, but the manuscript does not describe how the R2R recovery archive is constructed (behavior-descriptor definitions, stall threshold, archive size, or how the 'separate R2R archive briefly supplies' a policy is controlled), and no variance is reported. With only a +2.1 SR and −0.16 NE difference on an 11-scene subset, and no statistical support, this pilot does not currently substantiate the cross-domain transfer claim. Please either provide a full protocol and significance testing or explicitly recharacterize the pilot as illustrative rather than evidentiary.
minor comments (4)
  1. [Section 4.1] The text says 'Further experimental details are provided in the Appendix,' but no appendix is present in the manuscript; please include it or state the missing hyperparameters (population size, p_a, p_m, p_r, α, iteration budget T, archive size) in the main text.
  2. [Figure 3] The caption does not explain the difference between panels (a) and (b), and the behavior-grid axes are not labeled; please clarify what each panel shows.
  3. [Section 3.3.1] The behavior descriptors in Eqs. (4)–(5) depend on action categories such as 'interaction-oriented,' 'milestone-advancing,' and 'repeated low-value actions,' but the manuscript does not define how these categories are determined from raw trajectories; please provide the classification rule.
  4. [Section 4.1] The phrase 'standard food and stuff splits' is used without a definition or citation; please specify the exact episode split so the evaluation can be reproduced.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the central claim is empirical and the method's definitions do not reduce to the reported results.

full rationale

The paper's central claim is an empirical comparison (Table 1) and an ablation (Table 4), not a formal derivation. The behavior descriptors (Eqs. 4-5) are hand-defined statistics of rollouts; they are used to index the archive and to select switching candidates, but no theorem claims that descriptor diversity causes success. The quality score Q (Eq. 10) is defined directly from success rate and efficiency and is used for elite replacement, which is a standard optimization setup, not a prediction. The offline archive construction (Algorithm 1) and online switching (Algorithm 2) are evaluated empirically against baselines; a negative result was possible, so the claim is not forced by construction. The paper cites external prior work (MAP-Elites, ReAct, Reflexion, etc.) for context, but does not rely on self-citations as load-bearing evidence. The only concerns are reporting gaps: Section 4.1 defers experimental details to an appendix not included, and the text does not state whether the rollouts used in Algorithm 1 are disjoint from the scored evaluation episodes. If they were not disjoint, the Table 1 gains would be training-set selection, but that would be a validation flaw, not circularity under the definitions here, since no equation makes the reported success equal to an input by construction.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claims are empirical, so the main commitments are modeling assumptions (behavior descriptors, stall detection, policy representation) and unexamined hyperparameters. No new physical or mathematical entities are introduced.

free parameters (4)
  • alpha (quality weight) = not reported
    Eq. (10) balances success rate and efficiency; its value is not given and no sensitivity is shown.
  • stall threshold delta_s = 10
    Set in Section 4.1; no sensitivity analysis.
  • behavior grid resolution = 10x10
    Discretization of the behavior space in Section 4.1; not varied.
  • selection/mutation probabilities (p_a, p_m, p_r) = not reported
    Algorithm 1 relies on these probabilities; values are not given in the text.
assumptions (3)
  • domain assumption The behavior descriptor space defined by interaction intensity and goal-directedness (Eqs. 4-5) is a meaningful proxy for execution style and a useful coordinate for switching.
    Section 3.3.1 introduces the descriptors without validation that they capture behaviors that matter for recovery. If they do not, farthest-distance switching in Eq. (17) may select unhelpful policies.
  • domain assumption Sustained non-advancement of progress state over delta_s steps is a reliable operational signal of behavioral lock-in.
    Eq. (15) defines stall as equality of progress state across a window; this assumes progress state is updated at milestone boundaries and that repeated interactions without progress are the dominant failure mode.
  • domain assumption The fixed task scaffold plus editable rule modules parameterization (Eq. 2) is expressive enough for the QD search to discover genuinely different effective behaviors.
    Section 3.3.1 restricts policies to four rule modules; if this representation cannot express the behavioral variety needed for recovery, the archive will not contain useful alternatives.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Discovering Diverse Planning Policies for Multimodal Embodied Agents with Quality-Diversity Optimization." pith.science (2026). https://pith.science/paper/JXKECGKH

@misc{pith2026260808523,
  author       = {Pith},
  title        = {Pith review of: Discovering Diverse Planning Policies for Multimodal Embodied Agents with Quality-Diversity Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JXKECGKH}},
  note         = {Machine review of arXiv:2608.08523}
}
read the original abstract

Multimodal embodied agents are increasingly required to solve long-horizon tasks by integrating visual observations, textual goals, and interaction history into closed-loop decision making. However, state-of-the-art large-model-based planners often rely on a single dominant planning style during execution. Once this execution mode becomes ineffective, the agent may remain stalled for many steps, repeatedly interacting with the environment without making meaningful progress. We address this limitation by proposing a Quality-Diversity (QD) framework for discovering diverse planning policies for multimodal embodied agents. The proposed method treats planning-policy templates as evolvable individuals and organizes them into a behavior-indexed archive rather than collapsing search to a single prompt style. In the offline stage, rollout trajectories are summarized into structured success and failure experiences, which guide policy variation through recombination and experience-guided mutation. The resulting policies are mapped into a behavior space defined by interaction intensity and goal-directedness, and the highest-quality policy in each niche is retained in the archive. In the online stage, the agent executes one policy at a time while monitoring task progress. When persistent stall is detected, the system rolls back to the latest checkpoint and switches to a behaviorally distinct archive policy to resume execution. Experiments on the ThreeDWorld transport benchmark show that the proposed framework improves both task success and interaction efficiency over representative baseline planners. These results suggest that discovering diverse policy repertoires is an effective way to support adaptive multimodal planning and online failure recovery.

Figures

Figures reproduced from arXiv: 2608.08523 by the authors.

Figure 1
Figure 1. Motivation of archive-based behavioral recovery. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed framework. The method contains an offline archive-construction stage and an online [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. QD behavior-grid visualization (10×10); colored cells indicate occupied elite niches across the interaction-intensity and goal-directedness descriptors. stalls under one behavior mode, the runtime can select an elite from a different region of the grid and resume progress under the same strict step budget. These results support the role of behavioral diversity in long￾horizon recovery: when execution stalls, the arc… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Case-study trajectories in a representative transport [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 12 canonical work pages

  1. [1]

    Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Haus- man, et al. 2022. Do As I Can, Not As I Say: Grounding Language in Robotic Affordances.arXiv preprint arXiv:2204.01691(2022)

    arXiv 2022

  2. [2]

    Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Soren S"onderby, Ian Reid, and Anton van den Hengel. 2018. Vision-and-Language Navigation: Interpreting Visually-Grounded Navigation Instructions in Real En- vironments. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 3674–3683

    work page 2018

  3. [3]

    Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. 2017. Matterport3D: Learning from RGB-D Data in Indoor Environments. In2017 International Con- ference on 3D Vision. IEEE, 667–676

    work page 2017

  4. [4]

    Shoubin Chen, Zehao Wu, Kai Zhang, Chunyu Li, Baiyang Zhang, Fei Ma, Fei Richard Yu, and Qingquan Li. 2025. Exploring embodied multimodal large models: Development, datasets, and future directions.Information Fusion122 (2025), 103198

    work page 2025

  5. [5]

    Chrisantha Fernando, Dylan Sunil Banarse, Henryk Michalewski, Simon Osindero, and Tim Rocktäschel. 2024. Promptbreeder: Self-Referential Self- Improvement via Prompt Evolution. InInternational Conference on Machine Learning. PMLR, 13481–13544

    2024

  6. [6]

    Chuang Gan, Jeremy Schwartz, Seth Alter, Damian Mrowca, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, et al. 2020. ThreeDWorld: A Platform for Interactive Multi-Modal Physical Simulation.arXiv preprint arXiv:2007.04954(2020)

    arXiv 2020

  7. [7]

    Chuang Gan, Siyuan Zhou, Jeremy Schwartz, Seth Alter, Abhishek Bhandwaldar, Dan Gutfreund, Daniel LK Yamins, James J DiCarlo, Josh McDermott, Antonio Torralba, et al. 2022. The ThreeDWorld Transport Challenge: A Visually Guided Task-and-Motion Planning Benchmark Towards Physically Realistic Embodied AI. In2022 International conference on robotics and auto...

    work page 2022

  8. [8]

    Junyu Gao, Xuan Yao, Yong Rui, and Changsheng Xu. 2025. Building Embodied EvoAgent: A Brain-Inspired Paradigm for Bridging Multimodal Large Models and World Models. InProceedings of the 33rd ACM International Conference on Multimedia. 3280–3289

    work page 2025

Show all 23 references
  1. [9]

    Junyu Gao, Xuan Yao, and Changsheng Xu. 2024. Fast-slow test-time adaptation for online vision-and-language navigation. InProceedings of the 41st International Conference on Machine Learning. 14902–14919

    2024

  2. [10]

    Qingyan Guo, Rui Wang, Junliang Guo, Bei Li, Kaitao Song, Xu Tan, Guoqing Liu, Jiang Bian, and Yujiu Yang. 2024. EvoPrompt: Connecting LLMs with Evolu- tionary Algorithms Yields Powerful Prompt Optimizers. InThe Twelfth Interna- tional Conference on Learning Representations. h...

    2024

  3. [11]

    Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, et al. 2023. Inner Monologue: Embodied Reasoning through Planning with Language Models. In Conference on Robot Learning. PMLR, 1769–1782

    2023

  4. [12]

    Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee

  5. [13]

    Joel Lehman and Kenneth O Stanley. 2011. Abandoning objectives: Evolution through the search for novelty alone.Evolutionary computation19, 2 (2011), 189–223

    2011

  6. [14]

    Jean-Baptiste Mouret and Jeff Clune. 2015. Illuminating search spaces by mapping elites.arXiv preprint arXiv:1504.04909(2015)

    2015 arXiv

  7. [15]

    Justin K Pugh, Lisa B Soros, and Kenneth O Stanley. 2016. Quality diversity: A new frontier for evolutionary computation.Frontiers in Robotics and AI3 (2016), 40

    2016

  8. [16]

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language agents with verbal reinforcement learning. Advances in neural information processing systems36 (2023), 8634–8652

    2023

  9. [17]

    Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Cote, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht. 2021. ALFWorld: Aligning Text and Em- bodied Environments for Interactive Learning. InInternational Conference on Learning Representations. https://openreview.net/forum?id=0...

    2021

  10. [18]

    Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. 2023. LLM-Planner: Few-Shot Grounded Planning for Embodied Agents with Large Language Models. InProceedings of the IEEE/CVF international conference on computer vision. 2998–3009

    2023

  11. [19]

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language models.Advances in neural information processing systems36 (2023), 11809–11822

    2023

  12. [20]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. InThe eleventh international conference on learning representations

    2022

  13. [21]

    Hongxin Zhang, Weihua Du, Jiaming Shan, Qinhong Zhou, Yilun Du, Joshua B Tenenbaum, Tianmin Shu, and Chuang Gan. 2023. Building cooperative embodied agents modularly with large language models.arXiv preprint arXiv:2307.02485 (2023)

    2023 arXiv

  14. [22]

    Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, et al. 2023. RT-2: Vision-Language- Action Models Transfer Web Knowledge to Robotic Control. InConference on Robot Learning. PMLR, 2165–2183

    2023

  15. [2020]

    InEuropean Conference on Computer Vision

    Beyond the Nav-Graph: Vision-and-Language Navigation in Continuous Environments. InEuropean Conference on Computer Vision. Springer, 104–120

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.