Pith. sign in

REVIEW 4 major objections 4 minor 47 references

A photorealistic 3D-Gaussian simulator shows vision-based social navigation is far from solved, with top baselines succeeding under 22% of easy episodes.

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 03:03 UTC pith:WFBH2IRH

load-bearing objection A genuinely new 3DGS-based social navigation benchmark with a strong zero-shot evaluation, but the real-world transfer evidence is too thin to support the paper's central sim-to-real claim. the 4 major comments →

arxiv 2607.25219 v1 pith:WFBH2IRH submitted 2026-07-28 cs.RO

SONG: A Photorealistic 3D Gaussian Simulation Platform for Benchmarking Social Navigation

Weiqi Huang , Dianyi Yang , Jiaxin Li , Shuangyi Dong , Hao Xu , Zan Wang , Wei Liang This is my paper
classification cs.RO
keywords social navigation3D Gaussian splattingsimulation benchmarkvision-based navigationpedestrian simulationsim-to-real transfercollision detection
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 aims to build a high-fidelity simulation testbed for vision-based social navigation, using 3D Gaussian Splatting to render both scenes and human avatars, LLM-driven semantic intents to generate pedestrian trajectories, and a motion generator to animate avatars. The authors curate SONG-Bench, 500 episodes stratified by difficulty, and show that four representative navigation policies succeed in under 22% of easy episodes and near zero in hard ones. They further find that safety failures (collisions) dominate before any social-etiquette considerations, and that training on real-world data helps more than model scale. Finally, they show that fine-tuning on data from the platform improves real-world success of a physical robot, from 10% to 20% in single-pedestrian and from 5% to 10% in two-pedestrian scenarios.

Core claim

Using photorealistic 3D Gaussian rendering of environments and animated human avatars, the platform reveals that state-of-the-art vision-based navigation policies cannot reliably navigate human-populated scenes: across 500 episodes, success rates remain below 22% even in the easiest tier, and near zero in the hardest. The dominant failure mode is collision—both with static obstacles and with avatar Gaussians—occurring more often than any social-comfort violation. The performance gap between baselines vanishes in avatar-free scenes, indicating that perception and avoidance of dynamic humans, not scene geometry, is the critical bottleneck. As evidence that the simulator's fidelity transfers, f

What carries the argument

3D Gaussian Splatting for both static scenes and animatable human avatars: every scene and pedestrian is represented as a set of Gaussian primitives, rendered in real time from the robot's onboard camera, while pedestrian trajectories come from an LLM reasoning over semantic maps, and full-body motion is synthesized by a trajectory-conditioned generator. Collisions with pedestrians are detected by counting Gaussian centers inside a fixed robot volume, with a threshold of five centers flagging contact.

Load-bearing premise

The avatar-collision metric counts Gaussian centers inside a fixed robot volume and flags collision when the count exceeds 5; because Gaussian density varies across avatars and body regions, this proxy may not match true body contact, and if it is wrong the reported safety rates and the 'safety before etiquette' conclusion would change.

What would settle it

One could directly measure physical contact in a controlled experiment: place a real human in the environments or scan the same human with a calibrated body model, run the same episodes, and compare the reported avatar-collision counts to ground-truth body contact; if the count threshold mispredicts contact in a substantial fraction of cases, the safety metrics and the ordering of baselines would need revision.

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

If this is right

  • Benchmarking on SONG-Bench provides a standardized, photorealistic testbed for future vision-based social navigation research.
  • The finding that safety deficits precede social etiquette suggests that improving basic collision avoidance and recovery should come before modeling social norms.
  • The advantage of real-world-trained data over model scale implies that simulator realism—not architecture—is a key lever for progress.
  • The real-robot transfer results indicate that high-fidelity synthetic data can be used to fine-tune policies for physical deployment.
  • The benchmark's deterministic episode caches enable reproducible comparisons across policies.

Where Pith is reading between the lines

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

  • The Gaussian-density collision proxy may conflate visual density with physical contact, so safety numbers should be interpreted as relative rankings rather than absolute collision counts.
  • The non-interactive pedestrian assumption limits conclusions to unilateral adaptation; reciprocal human-robot interaction remains a distinct challenge.
  • The platform could be extended to generate reactive pedestrians using learned response models, turning SONG-Bench into a two-way interaction benchmark.
  • Because avatar-free success rates are still low, part of the difficulty is visual grounding in photorealistic scenes, which may benefit from domain randomization or pretraining.

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 / 4 minor

Summary. SONG introduces a 3D Gaussian splatting simulation platform for vision-based social navigation, combining 1,000 3DGS scenes, 500 animatable Gaussian avatars, LLM-generated pedestrian intents, Kimodo-synthesized full-body motion, and a closed-loop Isaac Lab runtime. The authors curate SONG-Bench (500 episodes stratified into easy/medium/hard) and propose a metric suite covering effectiveness, safety, and social compliance. Zero-shot evaluation of four pretrained baselines (NoMaD, NavDP, iPlanner, SocialNav) shows success rates below 22% in easy settings and near zero in hard settings; an avatar-free ablation attributes much of the failure to static obstacle collisions; and a real-robot experiment on a Unitree Go2 reports that fine-tuning NavDP on SONG data raises success from 10% to 20% (single pedestrian) and from 5% to 10% (two pedestrians).

Significance. If the platform performs as described, it would fill a genuine gap: a photorealistic, dynamic, semantically meaningful social-navigation benchmark with closed-loop visual observation. The deterministic episode caches, the use of external pretrained baselines in zero-shot evaluation, human-teleoperation upper bounds, and the modular toolchain are concrete strengths. The headline findings are largely diagnostic, and the authors appropriately acknowledge some limitations. However, the two pieces of evidence that carry the strongest claims — the avatar-collision safety metric and the real-world transfer result — are either an acknowledged proxy or statistically underpowered, so the significance of the paper as a validation of sim-to-real transfer is presently conditional on additional evidence.

major comments (4)
  1. [§5.2, Table 4] The real-world transfer claim is not supported by the reported data. The improvement is 2/20 to 4/20 successes in the single-pedestrian scenario and 1/20 to 2/20 in the two-pedestrian scenario. A Fisher exact test gives one-sided p≈0.33 and p≈0.49, and the 95% confidence intervals for the success-rate difference include zero. No confidence intervals, effect sizes, or significance tests are reported. The conclusion that fine-tuning on SONG data 'effectively improves the success rate in real-world environments' therefore rests on sampling noise. Please report per-trial details, increase the number of trials with a pre-specified protocol, or explicitly reframe this as a preliminary trend rather than a validated improvement.
  2. [§3.4, Appendix C] The dynamic-avatar collision metric counts the number of Gaussian centers within a 0.2 m radius spherical volume and flags a collision when that count exceeds η_col=5. The authors concede in Appendix C that Gaussian density varies across avatars, clothing regions, and reconstruction sources, so a fixed threshold is only an approximation of body-level contact. This proxy directly feeds CR_hum and the 'safety deficit precedes social etiquette' conclusion. Please calibrate the threshold against a mesh/SMPL-X or SDF-based body contact check on a sample of avatars and episodes, and report the sensitivity of CR_hum and the safety conclusions to η_col and to per-avatar density statistics.
  3. [Table 3, §5.1] The avatar-free ablation is used to support the claim that 'real-world data matters more than model scale' and that the performance ranking in Table 2 stems from handling dynamic human agents. However, the four baselines differ in architecture, observation modality, and goal representation (e.g., NoMaD uses image goals; NavDP and iPlanner use coordinates/depth; SocialNav uses RGB plus text). The ablation therefore cannot separate training-data distribution from these confounds. Please add controls that vary data distribution while holding architecture and goal representation fixed, or at least temper the causal claims and report per-model failure-mode analysis.
  4. [§5.1, Appendix B, Table 2] There is a direct inconsistency in the time budget. Section 5.1 sets Tmax=200 steps, and Appendix B says 'each episode runs for at most 200 steps at 10 Hz', which gives a 20 s horizon. Yet Table 2 reports successful-episode AvgT values as high as 120.88 s (NavDP medium) and 92.41 s (NoMaD medium), which are impossible under that budget. Please clarify what Tmax actually measures (seconds? steps at what control frequency?) and correct the table, the protocol text, and any downstream timeout computations accordingly.
minor comments (4)
  1. [§4.1, Eq. (1)] The success condition requires C_hum=0, but PSC is computed with a 0.5 m nearest-avatar threshold. Please clarify whether an episode with a near-miss <0.5 m but no counted collision is successful yet penalized on PSC, as the current metric definitions imply.
  2. [References] Several reference entries contain typos or informal formatting (e.g., 'Mingcbio Sun' in [8]; 'V ogel' in [7]; 'Mononen Mikko and Recast Navigation contributors' with a webpage-style citation [31]). Please normalize the bibliography.
  3. [§5.1, NoMaD evaluation] The paper states that NoMaD receives the RGB image rendered at the goal position as its image goal. This gives NoMaD access to goal appearance rather than a spatial coordinate. Please discuss the fairness implications of this protocol, since it differs from the coordinate-based goal input used by the other baselines.
  4. [Appendix E] Table A2 reports full closed-loop runtime at 10.01 steps/s. With Tmax=200 steps, this would imply about 20 seconds per episode, which further highlights the need to reconcile the time-budget description.

Circularity Check

0 steps flagged

No circularity: zero-shot baselines use external pretrained weights, real-world transfer is measured on a physical robot, and no fitted input is renamed as a prediction.

full rationale

I walked the derivation chain: SONG-Bench's headline numbers come from running externally released pretrained baselines (NoMaD, NavDP, iPlanner, SocialNav) zero-shot in a 3DGS simulator, and the sim-to-real claim is tested on a physical Unitree Go2 with a RealSense camera. Neither result is an input to the paper's own construction: the baselines were not fitted to SONG, and the real-world success rates were measured outside the simulator. The difficulty stratification (Section 4.2) uses path length, turns, avatar encounters, and corridor width; the success criterion (Equation 1) includes C_hum=0, so harder episodes naturally offer more collision opportunities, but the specific success values are empirical measurements, not identities forced by definition. The real-world fine-tuning comparison (Table 4, 10%->20% and 5%->10%) is statistically underpowered, but that is a correctness/evidence concern, not circularity. The self-citations in Related Work are contextual and not load-bearing; no uniqueness theorem or ansatz is imported from the authors' prior work to force a conclusion. The Appendix C limitation about the Gaussian-center collision proxy is an acknowledged validity caveat, not a circular reduction: the safety metrics depend on an approximate contact model, but that model is an independent benchmark design choice rather than a fitted parameter renamed as a prediction. Overall, the paper's central measurements are self-contained against external benchmarks and no step reduces by construction to its own inputs.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

SONG introduces no physical entities, forces, or conserved quantities. The listed free parameters are hand-set benchmark thresholds and difficulty-scale choices that define the task and metrics. The central claims are empirical and depend on these choices plus assumptions about asset realism, behavioral plausibility, and the collision proxy.

free parameters (5)
  • Avatar collision threshold η_col = 5 Gaussian centers
    A human collision is declared when more than 5 avatar Gaussian centers lie inside the robot's spherical collision volume (Sec. 3.4, Appendix B). No calibration against physical contact is given, and density varies by avatar (Appendix C).
  • Goal/obstacle success thresholds = δ_goal=1.0 m; δ_obs=5; T_max=200 steps
    Equation (1) defines success using these hand-set thresholds; changing them changes SR and all downstream findings.
  • Personal-space compliance threshold = 0.5 m
    PSC counts steps where nearest-avatar distance is at least 0.5 m (Sec. 4.3). The threshold is not justified from human-robot interaction literature.
  • Difficulty binning weights = 4 factors, tertile scores 0-2; labels top 100/200/200
    The easy/medium/hard partition is built from equal-frequency bins on path length, turns, encounters, and corridor width. Difficulty labels and reported per-level success rates depend on this binning.
  • Pedestrian speed range = 0.5–1.0 m/s
    Task dynamics and episode difficulty depend on the sampled pedestrian speed range (Sec. 4.1).
axioms (6)
  • domain assumption 3DGS rendering of SAGE-3D scenes and HumanGS/LHM avatars is visually faithful enough to serve as a real-world proxy.
    The sim-to-real transfer claim depends on this; no perceptual-similarity metric is provided.
  • domain assumption LLM-generated intents plus Recast/DetourCrowd trajectories produce socially plausible pedestrian behavior.
    Behavioral plausibility is asserted but not quantitatively validated beyond qualitative snapshots.
  • domain assumption Kimodo motion plus post-processing produces natural full-body motion when bound to Gaussian avatars.
    No motion-naturalness evaluation is given; Appendix B mentions foot-ground correction but no human or automated quality metric.
  • domain assumption Fixed non-interactive pedestrian trajectories are sufficient to benchmark social navigation and draw conclusions about current model deficits.
    Acknowledged in Appendix C: pedestrians do not react to the robot, so reciprocal interaction is not captured.
  • domain assumption The four selected baselines with their public weights are representative SOTA methods and are fairly adapted to SONG.
    Zero-shot conclusions depend on baseline selection and adaptation choices, e.g. rendering the goal image for NoMaD rather than its native goal interface.
  • ad hoc to paper Counting Gaussian centers inside a spherical volume with threshold η_col=5 approximates body-level collision.
    The collision proxy is introduced by this paper and explicitly flagged in Appendix C as an approximation because Gaussian density varies across avatars.

pith-pipeline@v1.3.0-alltime-deepseek · 14676 in / 12149 out tokens · 119099 ms · 2026-08-01T03:03:32.820526+00:00 · methodology

0 comments
read the original abstract

Social navigation has progressed from simplified 2D environments toward a more general vision-based setting, in which a robot needs to achieve socially compliant behavior purely from onboard visual observations. Yet supporting simulation platforms have not kept pace: existing options either lack visual observations, lack moving human avatars, or fall short of real-world fidelity in appearance and pedestrian behavior, offering limited support for advancing vision-based social navigation. We introduce SONG, a SOcial Navigation platform powered by 3D Gaussian splatting (3DGS). It leverages 3DGS for both scene and avatar representations, drives pedestrians using semantically grounded trajectories generated by a large language model, and synthesizes their full-body motion with a trajectory-conditioned generator to produce continuous, natural movement. On top of the platform, we curate SONG-Bench, a set of evaluation episodes stratified by difficulty, and propose a multi-dimensional metric suite covering effectiveness, safety, and social compliance. A systematic evaluation of representative navigation baselines reveals three findings: (a) vision-based social navigation is far from solved; (b) a critical safety deficit precedes social etiquette; (c) real-world data matters more than model scale. Crucially, we demonstrate that fine-tuning on our curated data effectively improves the success rate in real-world environments. We hope our platform provides a faithful and rigorous testbed for the next generation of vision-based social navigation research.

Figures

Figures reproduced from arXiv: 2607.25219 by Dianyi Yang, Hao Xu, Jiaxin Li, Shuangyi Dong, Wei Liang, Weiqi Huang, Zan Wang.

Figure 1
Figure 1. Figure 1: SONG Overview. We introduce a high-fidelity benchmark for social navigation. (a) Our asset library comprises 1,000 photorealistic 3DGS environments and 500 3DGS dynamic human avatars. (b) The simulation engine generates reasonable social episodes with tiered complexity, providing multi-modal first-person perceptions. (c) A holistic evaluation framework benchmarks SOTA policies across effectiveness, safety,… view at source ↗
Figure 2
Figure 2. Figure 2: SONG Pipeline. The platform prepares Gaussian scene and avatar assets, synthesizes social episodes from semantic intents, trajectories, motions, and avatar bindings, and exposes the resulting dynamic Gaussian scene through a closed-loop runtime for rendering and collision-aware agent interaction. and rich benchmarking tooling, similarly relies on low-fidelity mesh avatars with rule-based motion, leaving th… view at source ↗
Figure 3
Figure 3. Figure 3: SONG-Bench Overview. Left: episode statistics across difficulty levels, measured by human avatar encounters and optimal path length. Middle: scene snapshots from SONG-Bench under easy, medium, and hard settings. Right: visual examples from other representative platforms, including Habitat-Crowd [42], Falcon [12], and Arena5.0 [19]. 4 Evaluation protocol 4.1 Task definition SONG-Bench focuses on the vision-… view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of navigation behaviors. generated by a fine-tuned Height model [28]. Specifically, we first fine-tune the Height planner on 2D semantic maps using 5k episodes across 500 diverse scenarios. During this stage, the model is provided with rich spatial priors, including the precise positions of all pedestrians and obstacles, achieving a 75% success rate on these 5k training episodes. We then sele… view at source ↗
Figure 5
Figure 5. Figure 5: Real-World Experiment Visualization. The left part illustrates the real-world robotic platform and sensor setup, while the right part shows navigation trajectories and collision outcomes of NavDP under fine-tuned and zero-shot settings in one-person and two-person walking scenarios. Real-world experimental setup. As shown in [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

47 extracted references · 8 linked inside Pith

  1. [1]

    On evaluation of embodied navigation agents.arXiv preprint arXiv:1807.06757, 2018

    Peter Anderson, Angel Chang, Devendra Singh Chaplot, Alexey Dosovitskiy, Saurabh Gupta, Vladlen Koltun, Jana Kosecka, Jitendra Malik, Roozbeh Mottaghi, Manolis Savva, et al. On evaluation of embodied navigation agents.arXiv preprint arXiv:1807.06757, 2018. 1, 7

    Pith/arXiv arXiv 2018

  2. [2]

    Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments

    Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. InConference on Computer Vision and Pattern Recognition (CVPR), 2018. 2

    2018

  3. [3]

    Soc- navbench: A grounded simulation testing framework for evaluating social navigation.J

    Abhijat Biswas, Allan Wang, Gustavo Silvera, Aaron Steinfeld, and Henny Admoni. Soc- navbench: A grounded simulation testing framework for evaluating social navigation.J. Hum.- Robot Interact., 11(3):24, 2022. 2, 3

    2022

  4. [4]

    Navdp: Learning sim-to-real navigation diffusion policy with privileged information guidance.arXiv preprint arXiv:2505.08712, 2025

    Wenzhe Cai, Jiaqi Peng, Yuqiang Yang, Yujian Zhang, Meng Wei, Hanqing Wang, Yilun Chen, Tai Wang, and Jiangmiao Pang. Navdp: Learning sim-to-real navigation diffusion policy with privileged information guidance.arXiv preprint arXiv:2505.08712, 2025. 7, 8

    arXiv 2025

  5. [5]

    Object goal navigation using goal-oriented semantic exploration

    Devendra Singh Chaplot, Dhiraj Gandhi, Abhinav Gupta, and Ruslan Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. InAdvances in Neural Information Processing Systems (NeurIPS), 2020. 1

    2020

  6. [6]

    Crowd-robot interaction: Crowd-aware robot navigation with attention-based deep reinforcement learning

    Changan Chen, Yuejiang Liu, Sven Kreiss, and Alexandre Alahi. Crowd-robot interaction: Crowd-aware robot navigation with attention-based deep reinforcement learning. InInterna- tional Conference on Robotics and Automation (ICRA), 2019. 2, 3

    2019

  7. [7]

    F3loc: Fusion and filtering for floorplan localization

    Changan Chen, Rui Wang, Christoph V ogel, and Marc Pollefeys. F3loc: Fusion and filtering for floorplan localization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18029–18038, 2024. 2

    2024

  8. [8]

    Socialnav: Training human-inspired foundation model for socially-aware embodied navigation, 2025c.arXiv preprint arXiv:2511.21135, 2025

    Ziyi Chen, Yingnan Guo, Zedong Chu, Minghua Luo, Yanfen Shen, Mingcbio Sun, Junjun Hu, Shichao Xie, Kuan Yang, Pei Shi, et al. Socialnav: Training human-inspired foundation model for socially-aware embodied navigation, 2025c.arXiv preprint arXiv:2511.21135, 2025. 2, 3, 7, 8

    arXiv 2025

  9. [9]

    Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering

    Wei Cheng, Ruixiang Chen, Wanqi Yin, Siming Fan, Keyu Chen, Honglin He, Huiwen Luo, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, Daxuan Ren, Lei Yang, Ziwei Liu, Chen Change Loy, Chen Qian, Wayne Wu, Dahua Lin, Bo Dai, and Kwan-Yee Lin. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. arXiv pre...

    Pith/arXiv arXiv 2023

  10. [10]

    The dynamic window approach to collision avoidance.IEEE robotics & automation magazine, 4(1):23–33, 2002

    Dieter Fox, Wolfram Burgard, and Sebastian Thrun. The dynamic window approach to collision avoidance.IEEE robotics & automation magazine, 4(1):23–33, 2002. 3

    2002

  11. [11]

    Principles and guidelines for evaluating social robot navigation algorithms.ACM Transactions on Human- Robot Interaction, 14(2):1–65, 2025

    Anthony Francis, Claudia Pérez-d’Arpino, Chengshu Li, Fei Xia, Alexandre Alahi, Rachid Alami, Aniket Bera, Abhijat Biswas, Joydeep Biswas, Rohan Chandra, et al. Principles and guidelines for evaluating social robot navigation algorithms.ACM Transactions on Human- Robot Interaction, 14(2):1–65, 2025. 2

    2025

  12. [12]

    From cognition to precognition: A future-aware framework for social navigation

    Zeying Gong, Tianshuai Hu, Ronghe Qiu, and Junwei Liang. From cognition to precognition: A future-aware framework for social navigation. InInternational Conference on Robotics and Automation (ICRA), 2025. 2, 3, 6

    2025

  13. [13]

    Han, Hugues Thomas, Jian Zhang, Nicholas Rhinehart, and Timothy D

    James R. Han, Hugues Thomas, Jian Zhang, Nicholas Rhinehart, and Timothy D. Barfoot. Dr-mpc: Deep residual model predictive control for real-world social navigation.IEEE Robotics and Automation Letters (RA-L), 10(4):4029–4036, 2025. 2, 3

    2025

  14. [14]

    Ratatouille: Imitation learning ingredients for real-world social robot navigation.arXiv preprint arXiv:2509.17204, 2025

    James R Han, Mithun Vanniasinghe, Hshmat Sahak, Nicholas Rhinehart, and Timothy D Barfoot. Ratatouille: Imitation learning ingredients for real-world social robot navigation.arXiv preprint arXiv:2509.17204, 2025. 2, 3

    arXiv 2025

  15. [15]

    Ir-sim: An open-source lightweight simulator for robot navigation, control, and learning

    Ruihua Han. Ir-sim: An open-source lightweight simulator for robot navigation, control, and learning. Online, 2026. 2, 3 10

    2026

  16. [16]

    Social force model for pedestrian dynamics.Physical review E, 51(5):4282, 1995

    Dirk Helbing and Peter Molnar. Social force model for pedestrian dynamics.Physical review E, 51(5):4282, 1995. 3

    1995

  17. [17]

    Navthinker: Action-conditioned world models for coupled prediction and planning in social navigation.arXiv preprint arXiv:2603.15359, 2026

    Tianshuai Hu, Zeying Gong, Lingdong Kong, XiaoDong Mei, Yiyi Ding, Qi Zeng, Ao Liang, Rong Li, Yangyi Zhong, and Junwei Liang. Navthinker: Action-conditioned world models for coupled prediction and planning in social navigation.arXiv preprint arXiv:2603.15359, 2026. 2, 3

    arXiv 2026

  18. [18]

    Floor plan-guided visual navigation incorporating depth and directional cues.arXiv preprint arXiv:2511.01493,

    Weiqi Huang, Jiaxin Li, Zan Wang, Huijun Di, Wei Liang, and Zhu Yang. Floor plan-guided visual navigation incorporating depth and directional cues.arXiv preprint arXiv:2511.01493,

    arXiv

  19. [19]

    Arena 5.0: A photorealistic ros2 simulation framework for developing and benchmarking social navigation

    Linh Kästner. Arena 5.0: A photorealistic ros2 simulation framework for developing and benchmarking social navigation. InRobotics: Science and Systems (RSS), 2025. 3, 6

    2025

  20. [20]

    Arena 3.0: Advancing social navigation in collaborative and highly dynamic environments

    Linh Kästner, V olodymyir Shcherbyna, Huajian Zeng, Tuan Anh Le, Maximilian Ho-Kyoung Schreff, Halid Osmaev, Nam Truong Tran, Diego Diaz, Jan Golebiowski, Harold Soh, et al. Arena 3.0: Advancing social navigation in collaborative and highly dynamic environments. arXiv preprint arXiv:2406.00837, 2024. 3

    Pith/arXiv arXiv 2024

  21. [21]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. InACM Transactions on Graphics (SIGGRAPH),

  22. [22]

    Navigating to objects specified by images

    Jacob Krantz, Theophile Gervet, Karmesh Yadav, Austin Wang, Chris Paxton, Roozbeh Mot- taghi, Dhruv Batra, Jitendra Malik, Stefan Lee, and Devendra Singh Chaplot. Navigating to objects specified by images. InInternational Conference on Computer Vision (ICCV), 2023. 1

    2023

  23. [23]

    Beyond the nav- graph: Vision-and-language navigation in continuous environments

    Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav- graph: Vision-and-language navigation in continuous environments. InEuropean Conference on Computer Vision (ECCV), 2020. 2

    2020

  24. [24]

    C. Li, J. Jang, F. Xia, R. Martín-Martín, C. D’Arpino, A. Toshev, A. Francis, E. Lee, and S. Savarese. igibson challenge 2021. Online, 2021. 3, 7

    2021

  25. [25]

    Flona: Floor plan guided embodied visual navigation

    Jiaxin Li, Weiqi Huang, Zan Wang, Wei Liang, Huijun Di, and Feng Liu. Flona: Floor plan guided embodied visual navigation. InAAAI Conference on Artificial Intelligence (AAAI), 2025. 2

    2025

  26. [26]

    Intention aware robot crowd navigation with attention-based interaction graph.arXiv preprint arXiv:2203.01821,

    Shuijing Liu, Peixin Chang, Zhe Huang, Neeloy Chakraborty, Kaiwen Hong, Weihang Liang, D Livingston McPherson, Junyi Geng, and Katherine Driggs-Campbell. Intention aware robot crowd navigation with attention-based interaction graph.arXiv preprint arXiv:2203.01821,

    Pith/arXiv arXiv

  27. [27]

    Decentralized structural-rnn for robot crowd navigation with deep reinforcement learning

    Shuijing Liu, Peixin Chang, Weihang Liang, Neeloy Chakraborty, and Katherine Driggs- Campbell. Decentralized structural-rnn for robot crowd navigation with deep reinforcement learning. InInternational Conference on Robotics and Automation (ICRA), 2021. 2, 3

    2021

  28. [28]

    Shuijing Liu, Haochen Xia, Fatemeh Cheraghi Pouria, Kaiwen Hong, Neeloy Chakraborty, Zichao Hu, Joydeep Biswas, and Katherine Driggs-Campbell. Height: Heterogeneous interac- tion graph transformer for robot navigation in crowded and constrained environments.IEEE Transactions on Automation Science and Engineering, 23:1211–1230, 2025. 2, 3, 9

    2025

  29. [29]

    Core challenges of social robot navigation: A survey.ACM Transactions on Human-Robot Interaction, 12(3):1–39, 2023

    Christoforos Mavrogiannis, Francesca Baldini, Allan Wang, Dapeng Zhao, Pete Trautman, Aaron Steinfeld, and Jean Oh. Core challenges of social robot navigation: A survey.ACM Transactions on Human-Robot Interaction, 12(3):1–39, 2023. 3

    2023

  30. [30]

    Towards physically executable 3d gaussian for embodied navigation.arXiv preprint arXiv:2510.21307, 2025

    Bingchen Miao, Rong Wei, Zhiqi Ge, Xiaoquan Sun, Shiqi Gao, Jingzhe Zhu, Renhan Wang, Siliang Tang, Jun Xiao, Rui Tang, and Juncheng Li. Towards physically executable 3d gaussian for embodied navigation.arXiv preprint arXiv:2510.21307, 2025. 2, 4, 14

    arXiv 2025

  31. [31]

    Recast navigation, 2026

    Mononen Mikko and Recast Navigation contributors. Recast navigation, 2026. 5 11

    2026

  32. [32]

    Isaac lab: A gpu-accelerated simulation framework for multi-modal robot learning.arXiv preprint arXiv:2511.04831, 2025

    Mayank Mittal, Pascal Roth, James Tigue, Antoine Richard, Octi Zhang, Peter Du, Antonio Serrano-Muñoz, Xinjie Yao, René Zurbrügg, Nikita Rudin, et al. Isaac lab: A gpu-accelerated simulation framework for multi-modal robot learning.arXiv preprint arXiv:2511.04831, 2025. 5

    Pith/arXiv arXiv 2025

  33. [33]

    Human gaussian splatting: Real-time rendering of animatable avatars.arXiv preprint arXiv:2311.17113, 2024

    Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez- Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars.arXiv preprint arXiv:2311.17113, 2024. 4

    Pith/arXiv arXiv 2024

  34. [34]

    Habitat 3.0: A co-habitat for humans, avatars and robots.arXiv preprint arXiv:2310.13724, 2023

    Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander William Clegg, Michal Hlavac, So Yeon Min, et al. Habitat 3.0: A co-habitat for humans, avatars and robots.arXiv preprint arXiv:2310.13724, 2023. 3

    Pith/arXiv arXiv 2023

  35. [35]

    Lhm: Large animatable human reconstruction model from a single image in seconds.arXiv preprint arXiv:2503.10625, 2025

    Lingteng Qiu, Xiaodong Gu, Peihao Li, Qi Zuo, Weichao Shen, Junfei Zhang, Kejie Qiu, Weihao Yuan, Guanying Chen, Zilong Dong, and Liefeng Bo. Lhm: Large animatable human reconstruction model from a single image in seconds.arXiv preprint arXiv:2503.10625, 2025. 2, 5

    Pith/arXiv arXiv 2025

  36. [36]

    Kimodo: Generating 3d human motion in 3d scenes from text instructions.arXiv preprint arXiv:2603.19959, 2026

    Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Kimodo: Generating 3d human motion in 3d scenes from text instructions.arXiv preprint arXiv:2603.19959, 2026. 2, 5

    arXiv 2026

  37. [37]

    Habitat: A platform for embodied ai research

    Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. InInternational Conference on Computer Vision (ICCV), 2019. 2

    2019

  38. [38]

    Arena 4.0: A comprehensive ros2 development and benchmarking platform for human-centric navigation using generative-model-based environment generation

    V olodymyr Shcherbyna, Linh Kastner, Diego Diaz, Huu Giang Nguyen, Maximilian Ho-Kyoung Schreff, Tim Seeger, Jonas Kreutz, Ahmed Martban, Zhengcheng Shen, Huajian Zeng, et al. Arena 4.0: A comprehensive ros2 development and benchmarking platform for human-centric navigation using generative-model-based environment generation. InInternational Conference on...

    2025

  39. [39]

    igibson 1.0: A simulation environment for interactive tasks in large realistic scenes

    Bokui Shen, Fei Xia, Chengshu Li, Roberto Martín-Martín, Linxi Fan, Guanzhi Wang, Claudia Pérez-D’Arpino, Shyamal Buch, Sanjana Srivastava, Lyne Tchapmi, et al. igibson 1.0: A simulation environment for interactive tasks in large realistic scenes. InInternational Conference on Intelligent Robots and Systems (IROS), 2021. 2, 3

    2021

  40. [40]

    Nomad: Goal masked diffusion policies for navigation and exploration

    Ajay Sridhar, Dhruv Shah, Catherine Glossop, and Sergey Levine. Nomad: Goal masked diffusion policies for navigation and exploration. InInternational Conference on Robotics and Automation (ICRA), 2024. 7, 8

    2024

  41. [41]

    Reciprocal n-body collision avoidance

    Jur Van Den Berg, Stephen J Guy, Ming Lin, and Dinesh Manocha. Reciprocal n-body collision avoidance. InRobotics research: the 14th international symposium ISRR, 2011. 3

    2011

  42. [42]

    Habicrowd: A high performance simulator for crowd-aware visual navigation

    An Vuong, Toan Nguyen, Minh Nhat Vu, Baoru Huang, Huynh Thi Thanh Binh, Thieu V o, and Anh Nguyen. Habicrowd: A high performance simulator for crowd-aware visual navigation. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5821–5827. IEEE, 2024. 3, 6

    2024

  43. [43]

    Towards versatile embodied navigation

    Hanqing Wang, Wei Liang, Luc Van Gool, and Wenguan Wang. Towards versatile embodied navigation. InAdvances in Neural Information Processing Systems (NeurIPS), 2022. 1

    2022

  44. [44]

    Rethinking the embodied gap in vision-and-language navigation: A holistic study of physical and visual disparities

    Liuyi Wang, Xinyuan Xia, Hui Zhao, Hanqing Wang, Tai Wang, Yilun Chen, Chengju Liu, Qijun Chen, and Jiangmiao Pang. Rethinking the embodied gap in vision-and-language navigation: A holistic study of physical and visual disparities. InInternational Conference on Computer Vision (ICCV), 2025. 2

    2025

  45. [45]

    Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames

    Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. InInternational Conference on Learning Representations (ICLR), 2020. 1

    2020

  46. [46]

    iPlanner: Imperative Path Planning

    Fan Yang, Chen Wang, Cesar Cadena, and Marco Hutter. iPlanner: Imperative Path Planning. InProceedings of Robotics: Science and Systems, 2023. 7, 8 12

    2023

  47. [47]

    A person is casually walking forward at a natural pace with minimal waist sway

    Ziyu Zhu, Xilin Wang, Yixuan Li, Zhuofan Zhang, Xiaojian Ma, Yixin Chen, Baoxiong Jia, Wei Liang, Qian Yu, Zhidong Deng, et al. Move to understand a 3d scene: Bridging visual grounding and exploration for efficient and versatile embodied navigation. InInternational Conference on Computer Vision (ICCV), 2025. 2 13 Appendix A Assets visualization Asset exam...

    2025