Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

Short high-level UAV commands need multi-stage process and safety tests that goal-only route benchmarks miss.

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 · grok-4.5

2026-07-13 21:54 UTC pith:DCVX2U2E

load-bearing objection Solid diagnostic UAV HL-VLA benchmark with a real 3DGS–Mesh twin and process/safety metrics; the main soft spot is that TCR scores scripted path coverage more than open-ended semantic completion. the 4 major comments →

arxiv 2603.19822 v4 pith:DCVX2U2E submitted 2026-03-20 cs.CV

HUGE-Bench: A Benchmark for High-Level UAV Vision-Language-Action Tasks

Jingyu Guo , Ziye Chen , Ziwen Li , Zhengqing Gao , Jiaxin Huang , Hanlue Zhang , Fengming Huang , Yu Yao
show 2 more authors
Tongliang Liu Mingming Gong
This is my paper
classification cs.CV
keywords High-Level VLAUAV benchmarkvision-language-action3D Gaussian Splattingprocess-oriented metricscollision-aware evaluationdigital twin
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.

Real UAV operators usually give brief, high-intent orders rather than long step-by-step routes, yet most aerial vision-language benchmarks still score agents mainly on whether they reach a final point. This paper argues that those benchmarks therefore fail to diagnose the harder skills real missions need: grounding short language into multi-stage procedures, maintaining process fidelity, and avoiding collisions. HUGE-Bench supplies that missing testbed. It rebuilds four real outdoor scenes as aligned 3D Gaussian Splatting plus mesh digital twins, defines eight high-level tasks (landing, road and building inspection, mapping, orbiting, spiral descent, obstacle-aware traversal), and generates millions of meters of multimodal trajectories. Process coverage, terminal success, and collision-aware efficiency metrics then show that leading vision-language-action models leave large gaps on both semantic completion and safe execution. A sympathetic reader cares because the work reframes evaluation around the language and safety demands of actual UAV operations, not just endpoint navigation.

Core claim

Existing UAV vision-language benchmarks are poorly diagnostic for real operations because they emphasize long step-wise routes and goal-centric scores; when short high-level commands must be turned into multi-stage process-oriented trajectories with collision awareness, representative state-of-the-art VLA models show substantial gaps in semantic completion and safe execution. HUGE-Bench is presented as the diagnostic testbed that exposes those gaps.

What carries the argument

Aligned 3DGS–Mesh digital twin: photorealistic 3D Gaussian Splatting for perception paired with a metric mesh for collision queries and depth, enabling scalable rule-based trajectory generation and collision-aware scoring (TCR, SR, CR, CSPL) under short high-level instructions.

Load-bearing premise

Rule-generated trajectories scored by fixed distance thresholds and mesh collision flags in a static digital twin are a faithful enough proxy that model gaps will transfer to real high-level UAV process fidelity and safety.

What would settle it

Take the same short high-level instructions and run the same finetuned models on a physical UAV in the four real scenes (or a high-fidelity dynamic twin with moving obstacles); if process coverage and collision rates close the reported gaps, the claim that current VLA models fail high-level semantic completion and safety under this regime is undermined.

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

If this is right

  • Benchmark progress on UAV language control must report multi-stage process fidelity, not only endpoint success.
  • Collision-aware mesh geometry becomes a required evaluation channel alongside photorealistic rendering.
  • Short underspecified commands become the standard instruction style for high-level aerial VLA tests.
  • Models pretrained on large robot data transfer partially to aerial multi-stage tasks but still leave large residual errors on inspection, spiral, and traversal.
  • Future systems need explicit stage memory and safety planning to close the gaps the benchmark measures.

Where Pith is reading between the lines

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

  • The same short-command, process-plus-safety template could diagnose ground or underwater robots that also receive brief operator intent rather than route scripts.
  • Dynamic elements (moving people, weather, changing light) are the natural next stress test once static twins no longer separate models.
  • Stage annotations already present in the dataset invite hierarchical planners that predict subtask sequences before low-level actions.
  • If process metrics remain hard while endpoint success rises, public leaderboards may overstate readiness for inspection and search missions.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. HUGE-Bench proposes a UAV Vision–Language–Action benchmark for short, high-level commands that must be grounded into multi-stage, safety-aware flight rather than long route-following VLN. The resource comprises four real-world outdoor scenes reconstructed as aligned 3DGS–Mesh digital twins, eight HL-VLA tasks (landing, road/building inspection, mapping, orbiting, spiral-down, obstacle-aware traversal), and 2.56M meters of trajectories generated by task rules (and RRT for traversal) in Isaac Sim. Evaluation centers on Trajectory Coverage Rate (TCR@1/2/5) for process-oriented tasks, Success Rate for landing, and Collision Rate / Collision-aware SPL for traversal. Finetuned OpenVLA, FastVLM, π0, and π0.5 show large gaps on unseen splits, especially on multi-stage and safety-critical tasks, which the authors present as evidence that current VLA systems are not yet diagnostic-ready for high-level UAV autonomy.

Significance. If the evaluation protocol is accepted as measuring high-level process fidelity and safety, the paper fills a clear gap between route-following aerial VLN and operational short-command UAV control. Strengths include a real-to-sim capture pipeline, an aligned 3DGS–Mesh twin that jointly supports photorealistic perception and mesh collision queries, explicit train/seen/unseen splits with landmark- and instruction-level generalization, and head-to-head results across multiple SOTA VLA families. The hybrid representation and collision-aware metrics are practically useful for the community even if some metric interpretations need tightening. The work is a solid systems/benchmark contribution rather than a new learning method, and its value hinges on whether TCR and related scores truly diagnose semantic multi-stage completion rather than imitation of a single scripted path.

major comments (4)
  1. [§3.5 Evaluation Metrics; Table 2] §3.5 Eqs. (2)–(3) and Table 2: TCR defines process fidelity as coverage of one rule-generated ground-truth polyline (task-specific waypoints / RRT; §3.2). For Inspection-B/R, Mapping, Orbit-H/R, and Spiral Down, many geometrically different executions can satisfy the same high-level instruction (orbit direction, clearance, coverage order, spiral pitch). Low TCR can therefore mix failed semantic grounding with style mismatch to the generator, while high TCR can reward path cloning. This is load-bearing for the abstract/conclusion claim of “significant gaps in high-level semantic completion.” Please either (i) add multi-reference / stage-level metrics using the collected subtask phase labels (Fig. 2; Conclusion), (ii) report human or rule-based semantic process checks orthogonal to a single GT path, or (iii) reframe claims as imitation of scripted high-level procedures under short language
  2. [§3.2; §3.5; §4.2; Conclusion] The paper repeatedly advertises stage annotations and multi-stage decomposition (Fig. 2; Fig. 3; Conclusion: “explicit stage annotations and process-oriented metrics”), yet the reported protocol never evaluates stage success, stage order, or stage transition accuracy—only geometric TCR/SR/CR/CSPL. Given that stage labels are already logged, adding stage-completion rates (or failure breakdowns by stage) is necessary to support the HL-VLA diagnostic narrative and to separate early grounding failures from late execution drift.
  3. [Table 3; §4.2] Table 3 reports CR/CSPL only for FastVLM and π0 on Traversal; OpenVLA and π0.5 are omitted, and no other task receives collision analysis despite mesh collision signals being available throughout. Safety is a core claimed axis of HUGE-Bench (Abstract; Table 1; contributions). Please expand collision-aware evaluation to all baselines on Traversal and, where applicable, report en-route collisions for inspection/orbit tasks that enforce clearance, or justify why only two models and one task are safety-scored.
  4. [§4.2; Table 2; §3.5] §4.2 / Table 2: all primary numbers are single-run point estimates with no seeds, confidence intervals, or sensitivity to free parameters (TCR δ, landing success radius of 10 m, action horizon/execution window). Landing SR within 10 m is also a loose terminal criterion for a “landing” task and is not ablated. For a benchmark paper whose central product is comparative ranking and “significant gaps,” at least multi-seed variance on key tasks and a short sensitivity study on δ and the landing threshold are needed before the difficulty ordering and model ranking can be treated as stable.
minor comments (6)
  1. [Table 1] Table 1 “Len. (m/w)” is useful but the caption should define the instruction-length unit (words vs. tokens) and how multi-stage instructions are counted for HUGE-Bench vs. VLN baselines.
  2. [§3.1; §4.1] §3.1 Eq. (1) writes at = πθ(ot, x, st) but later baselines use two-frame visual inputs; clarify the observation definition used in experiments.
  3. [§3.3; Fig. 5] Fig. 5 word clouds and task proportions are informative; please report absolute trajectory counts per task and per scene, not only percentages.
  4. [Title; Abstract; §1] Minor typos/formatting: “UA V” in the title line; “cbrief”/“safecmulti-stage” artifacts in the abstract PDF text; inconsistent π0 / π 0.5 spacing; “dimen-sions” line break in contributions.
  5. [§2.3] Related work could more explicitly contrast process metrics with instruction-fidelity/coverage measures already used in VLN (cited [13]) to position TCR against prior alternatives.
  6. [Supplementary A–B; §4] Appendix real2sim vs sim2sim (Fig. 9) and finetune-vs-scratch (Fig. 10) strengthen the paper; consider promoting a short summary of those ablations into the main experiments section.

Circularity Check

0 steps flagged

No circular derivation: HUGE-Bench is an empirical benchmark; TCR/SR/CR/CSPL score independent model rollouts against rule-defined references, not a first-principles result reduced to its inputs.

full rationale

This paper does not claim a derivation, uniqueness theorem, or fitted-then-predicted quantity. It constructs a dataset (real-to-sim 3DGS–Mesh twins, task-rule/RRT waypoints, LLM spatial phrases), defines evaluation metrics (TCR as fraction of GT points within δ of the predicted polyline; SR; CR; CSPL), and reports performance of external VLA models (OpenVLA, FastVLM, π0, π0.5) on held-out seen/unseen splits. TCR (Eqs. 2–3) is explicitly a coverage score relative to a chosen reference trajectory—the same structure as classic VLN path-fidelity metrics (nDTW/SDTW/SPL)—not a prediction that reduces by construction to a fitted parameter. Training and evaluation share the digital twin, which is standard for simulation benchmarks and does not make reported gaps tautological: the policies are not the generators of the GT, and low TCR/CR is an empirical outcome under a fixed protocol. Concerns that fixed-rule GT under-represents multi-valid process semantics, or that sim gaps may not transfer to real flight, are validity/proxy issues, not circularity of a derivation chain. No self-definitional loop, no fitted input renamed as prediction, no load-bearing uniqueness imported from overlapping authors. Score 0.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 4 invented entities

The central claim rests on engineering and evaluation choices, not on new physics. Load-bearing premises are that short operator-style language plus multi-stage process metrics better match real UAV ops than long VLN routes; that mesh collisions and fixed distance thresholds operationalize safety and process fidelity; and that fine-tuned open VLA models under a two-frame protocol fairly expose capability gaps. Free parameters are metric thresholds and training knobs. Invented entities are the benchmark constructs (HL-VLA suite, TCR, CSPL, 3DGS–Mesh twin packaging), not new natural kinds.

free parameters (4)
  • TCR distance tolerances δ = 1 m / 2 m / 5 m
    Process coverage counts a GT point covered if min distance to the predicted polyline is < δ; reported at 1 m, 2 m, 5 m (§3.5). Choice of δ directly changes reported process fidelity.
  • Landing success radius = 10 m
    Landing SR uses a 10 m endpoint tolerance (§4.2), a hand-chosen success criterion that gates the terminal-accuracy claim.
  • Action horizon and execution window for π-family / FastVLM = horizon 20; exec first 10; 10 denoising steps
    Action horizon 20 with average of first 10 steps (π models) and 10 denoising steps (FastVLM) are training/inference knobs that affect measured performance (Appendix).
  • Trajectory sampling density = ~1 m / 5°
    Frames logged ~every 1 m and every 5° heading change during turns (§3.2); changes observation density and effective supervision.
axioms (5)
  • domain assumption Brief high-level natural-language commands plus multi-stage latent subtasks better represent real UAV operator practice than long step-wise VLN routes.
    Stated as the central mismatch in §1 and used to justify the entire HL-VLA formulation; not independently measured against field operator logs.
  • domain assumption Aligned mesh geometry collision queries are a sufficient safety oracle for evaluating UAV trajectories in the digital twin.
    §§1, 3.2, 3.5 treat mesh collisions as the safety channel for CR/CSPL; ignores dynamics, wind, sensor noise, and soft constraints.
  • ad hoc to paper Point-to-polyline coverage of a rule-generated reference trajectory (TCR) measures process fidelity of high-level semantic completion.
    Eqs. (2)–(3) define TCR; alternative valid processes (different orbit direction, coverage order) may be under-credited.
  • domain assumption Two-frame visual input (first frame + current) plus fine-tuning of existing VLA checkpoints is an adequate protocol to diagnose high-level UAV VLA capability.
    §4.1 follows prior work [30]; hierarchical planners or memory-heavy agents are not primary baselines.
  • standard math Standard imitation/flow-matching training and Euclidean path metrics are valid evaluation mathematics for continuous 3D flight policies.
    Uses conventional SR/SPL-style constructions and flow matching as in cited VLA work.
invented entities (4)
  • HL-VLA task suite (8 high-level UAV behaviors) no independent evidence
    purpose: Define the short-command, multi-stage evaluation regime distinct from route-following VLN.
    Tasks and subtask decompositions are specified by the authors in §3.1; independent operational standards are not cited as external ground truth.
  • Trajectory Coverage Rate (TCR) no independent evidence
    purpose: Score process fidelity of multi-stage trajectories beyond endpoint success.
    Introduced in §3.5 as a paper-specific metric; no external standard TCR for UAV HL tasks.
  • Collision-aware SPL (CSPL) no independent evidence
    purpose: Jointly score success, path efficiency, and collision-free execution.
    Defined in Eq. (4) by multiplying SPL-like terms by a no-collision indicator; paper-local combination of known ideas.
  • HUGE-Bench 3DGS–Mesh digital twin package no independent evidence
    purpose: Enable photoreal perception and collision-aware scalable data generation from real outdoor captures.
    Hybrid idea cites prior executable-3DGS work [21], but this specific four-scene corpus and pipeline are new artifacts of the paper.

pith-pipeline@v1.1.0-grok45 · 19316 in / 3932 out tokens · 40198 ms · 2026-07-13T21:54:38.120593+00:00 · methodology

0 comments
read the original abstract

Existing UAV vision-language navigation (VLN) benchmarks have enabled language-guided flight, but they largely focus on long, step-wise route descriptions with goal-centric evaluation, making them less diagnostic for real operations where brief, high-level commands must be grounded into safe multi-stage behaviors. We present HUGE-Bench, a benchmark for High-Level UAV Vision-Language-Action (HL-VLA) tasks that tests whether an agent can interpret concise language and execute complex, process-oriented trajectories with safety awareness. HUGE-Bench comprises 4 real-world digital twin scenes, 8 high-level tasks, and 2.56M meters of trajectories, and is built on an aligned 3D Gaussian Splatting (3DGS)-Mesh representation that combines photorealistic rendering with collision-capable geometry for scalable generation and collision-aware evaluation. We introduce process-oriented and collision-aware metrics to assess process fidelity, terminal accuracy, and safety. Experiments on representative state-of-the-art VLA models reveal significant gaps in high-level semantic completion and safe execution, highlighting HUGE-Bench as a diagnostic testbed for high-level UAV autonomy.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. MulRobBench: A Decision-Level Benchmark for Safe and Security-Policy-Compliant Multimodal UAV Agents

    cs.MA 2026-07 conditional novelty 6.0

    Across 17 multimodal models on 3,024 protocol-conditioned UAV decision samples, the best semantic protocol-decision score is only 0.5141 and strict mean dimension accuracy only 0.1599.

  2. Knowing the Self, Understanding the World: A Dual-Cognition Benchmark for UAV Spatio-temporal Reasoning with MLLMs

    cs.CV 2026-07 conditional novelty 6.0

    UAV-DualCog evaluates multimodal LLMs on self-state and environment-state reasoning in aerial images and videos, and shows current models are unreliable at spatial and temporal grounding.

Reference graph

Works this paper leans on

33 extracted references · 1 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Anderson, P., Chang, A.X., Chaplot, D.S., Dosovitskiy, A., Gupta, S., Koltun, V., Kosecka, J., Malik, J., Mottaghi, R., Savva, M., Zamir, A.: On evaluation of embodied navigation agents (2018)

    2018

  2. [2]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I., Gould, S., van den Hengel, A.: Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3674–3683 (2018)

    2018

  3. [3]

    In: International Conference on 3D Vision (3DV) (2017)

    Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor envi- ronments. In: International Conference on 3D Vision (3DV) (2017)

    2017

  4. [4]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2021)

    Chen, S., Guhur, P.L., Schmid, C., Laptev, I.: History aware multimodal trans- former for vision-and-language navigation. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)

    2021

  5. [5]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

    Chen, S., Li, Q., Qiu, H., Chen, Y., Wang, W., Xie, W., et al.: Think global, act lo- cal: Dual-scale graph transformer for vision-and-language navigation. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

    2022

  6. [6]

    Cheng,A.C.,Ji,Y.,Yang,Z.,Gongye,Z.,Zou,X.,Kautz,J.,Biyik,E.,Yin,H.,Liu, S., Wang, X.: Navila: Legged robot vision-language-action model for navigation (2024)

    2024

  7. [7]

    Gao, Y., Li, C., You, Z., Liu, J., Li, Z., Chen, P., Chen, Q., Tang, Z., Wang, L., Yang, P., Tang, Y., Tang, Y., Liang, S., Zhu, S., Xiong, Z., Su, Y., Ye, X., Li, J., Ding, Y., Wang, D., Wang, Z., Zhao, B., Li, X.: Openfly: A versatile toolchain and large-scale benchmark for aerial vision-language navigation (2025)

    2025

  8. [8]

    In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (ACL) (2022)

    Gu, J., Zhang, H., Zhao, H., Guo, J., et al.: Vision-and-language navigation: A survey of tasks, methods, and future directions. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (ACL) (2022)

    2022

  9. [9]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)

    Hao, W., Li, C., Li, X., Carin, L., Gao, J.: Towards learning a generic agent for vision-and-language navigation via pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)

    2020

  10. [10]

    Ilharco,G.,Jain,V.,Ku,A.,Ie,E.,Baldridge,J.:Generalevaluationforinstruction conditioned navigation using dynamic time warping (2019)

    2019

  11. [11]

    Intelligence, P., Black, K., Brown, N., Darpinian, J., Dhabalia, K., Driess, D., Esmail, A., Equi, M., Finn, C., Fusai, N., Galliker, M.Y., Ghosh, D., Groom, L., Hausman, K., Ichter, B., Jakubczak, S., Jones, T., Ke, L., LeBlanc, D., Levine, S., Li-Bell, A., Mothukuri, M., Nair, S., Pertsch, K., Ren, A.Z., Shi, L.X., Smith, L., Springenberg, J.T., Stachow...

    Pith/arXiv arXiv doi:10.48550/arxiv.2504.16054 2025

  12. [12]

    Intelligence, P., Black, K., Brown, N., Driess, D., Esmail, A., Equi, M., Finn, C., Fusai, N., Groom, L., Hausman, K., Ichter, B., Jakubczak, S., Jones, T., Ke, L., Levine, S., Li-Bell, A., Mothukuri, M., Nair, S., Pertsch, K., Shi, L.X., Tanner, J., Vuong, Q., Walling, A., Wang, H., Zhilinsky, U.:π0: A vision-language-action flow model for general robot ...

    doi:10.48550/arxiv.2410 2024

  13. [13]

    Guo et al

    Jain, V., Magalhaes, G., Ku, A., Vaswani, A., Ie, E., Baldridge, J.: Stay on the path: Instruction fidelity in vision-and-language navigation (2019) 16 J. Guo et al

    2019

  14. [14]

    In: 2011 IEEE international conference on robotics and automation

    Karaman, S., Walter, M.R., Perez, A., Frazzoli, E., Teller, S.: Anytime motion planning using the rrt. In: 2011 IEEE international conference on robotics and automation. pp. 1478–1483. ieee (2011)

    2011

  15. [15]

    ACM Transactions on Graphics42(4) (2023)

    Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics42(4) (2023)

    2023

  16. [16]

    Kim, M., Pertsch, K., Karamcheti, S., Xiao, T., Balakrishna, A., Nair, S., Rafailov, R., Foster, E., Lam, G., Sanketi, P., Vuong, Q., Kollar, T., Burchfiel, B., Tedrake, R., Sadigh, D., Levine, S., Liang, P., Finn, C.: Openvla: An open-source vision- language-action model (2024)

    2024

  17. [17]

    In: Computer Vision – ECCV 2020

    Krantz, J., Wijmans, E., Majumdar, A., Batra, D., Lee, S.: Beyond the nav-graph: Vision-and-language navigation in continuous environments. In: Computer Vision – ECCV 2020. pp. 104–120 (2020)

    2020

  18. [18]

    In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) (2020)

    Ku, A., Anderson, P., Patel, R., Ie, E., Baldridge, J.: Room-across-room: Multi- lingual vision-and-language navigation with dense spatiotemporal grounding. In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) (2020)

    2020

  19. [19]

    Lee, J., Miyanishi, T., Kurita, S., Sakamoto, K., Azuma, D., Matsuo, Y., Inoue, N.: Citynav: Language-goal aerial navigation dataset with geographic information (2024)

    2024

  20. [20]

    In: Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV)

    Liu, S., Zhang, H., Qi, Y., Wang, P., Zhang, Y., Wu, Q.: Aerialvln: Vision- and-language navigation for uavs. In: Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV). pp. 15338–15348 (2023).https: //doi.org/10.1109/ICCV51070.2023.01411

    doi:10.1109/iccv51070.2023.01411 2023

  21. [21]

    Miao, B., Wei, R., Ge, Z., Sun, X., Gao, S., Zhu, J., Wang, R., Tang, S., Xiao, J., Tang, R., Li, J.: Sage-3d: Towards physically executable 3d gaussian for embodied navigation (2025)

    2025

  22. [22]

    GitHub:https://github.com/isaac- sim/IsaacSim

    NVIDIA: NVIDIA Isaac Sim.https://docs.isaacsim.omniverse.nvidia.com/ latest/index.html(2025), version 5.1.0. GitHub:https://github.com/isaac- sim/IsaacSim. Accessed: 2026-03-06

    2025

  23. [23]

    In: International Conference on Learning Represen- tations (ICLR) (2024)

    Puig, X., Undersander, E., Szot, A., Dallaire Cote, M., Yang, T.Y., Partsey, R., Desai, R., Clegg, A.W., Hlavac, M., Min, S.Y., et al.: Habitat 3.0: A co-habitat for humans, avatars and robots. In: International Conference on Learning Represen- tations (ICLR) (2024)

    2024

  24. [24]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Qi, Y., Wu, Q., Anderson, P., Wang, X., Wang, W.Y., Shen, C., van den Hengel, A.: Reverie: Remote embodied visual referring expression in real indoor environments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9982–9991 (2020)

    2020

  25. [25]

    Ramakrishnan, S.K., Gokaslan, A., Wijmans, E., Maksymets, O., Clegg, A., Turner, J., Undersander, E., Galuba, W., Westbury, A., Chang, A., Savva, M., Zhao, Y., Batra, D.: Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai (2021),https://openreview.net/forum?id=- v4OuqNs5P, neurIPS Datasets and Benchmarks Track

    2021

  26. [26]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Savva, M., Kadian, A., Maksymets, O., Zhao, Y., Wijmans, E., Jain, B., Straub, J., Liu, J., Koltun, V., Malik, J., Parikh, D., Batra, D.: Habitat: A platform for embodied ai research. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9339–9347 (2019)

    2019

  27. [27]

    In: Field and Service Robotics (2017)

    Shah, S., Dey, D., Lovett, C., Kapoor, A.: Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In: Field and Service Robotics (2017)

    2017

  28. [28]

    In: Proceedings of the 37th International Conference on Machine Learning (ICML) (2020) HUGE-Bench 17

    Thomason, J., Murray, M., Cakmak, M., Zettlemoyer, L.: Vision-and-dialog navi- gation. In: Proceedings of the 37th International Conference on Machine Learning (ICML) (2020) HUGE-Bench 17

    2020

  29. [29]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Vasu, P.K.A., Faghri, F., Li, C., Koc, C., True, N., Antony, A., Santhanam, G., Gabriel, J., Grasch, P., Tuzel, O., Pouransari, H.: Fastvlm: Efficient vision encod- ing for vision language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19769–19780 (June 2025)

    2025

  30. [30]

    Wang, X., Yang, D., Liao, Y., Zheng, W., Wu, W., Dai, B., Li, H., Liu, S.: Uav- flow colosseo: A real-world benchmark for flying-on-a-word uav imitation learning (2025)

    2025

  31. [31]

    Wang, X., Yang, D., Wang, Z., Kwan, H., Chen, J., Wu, W., Li, H., Liao, Y., Liu, S.: Towards realistic uav vision-language navigation: Platform, benchmark, and methodology (2024)

    2024

  32. [32]

    In: Robotics: Science and Systems (RSS) (2025)

    Zhang, J., Wang, K., Wang, S., Li, M., Liu, H., Wei, S., Wang, Z., Zhang, Z., Wang, H.: Uni-navid: A video-based vision-language-action model for unifying embodied navigation tasks. In: Robotics: Science and Systems (RSS) (2025)

    2025

  33. [33]

    In: Robotics: Science and Systems (RSS) (2024) 18 J

    Zhang, J., Wang, K., Xu, R., Zhou, G., Hong, Y., Fang, X., Wu, Q., Zhang, Z., Wang, H.: Navid: Video-based vlm plans the next step for vision-and-language navigation. In: Robotics: Science and Systems (RSS) (2024) 18 J. Guo et al. Supplementary Material A Implementation Details 3DGS–Mesh Reconstruction from Real-World Data.We collect large- scale outdoor ...

    2024