Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

A single video of a room can be turned into a photorealistic training simulator that teaches a robot to avoid moving people in that same room.

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

2026-08-03 00:03 UTC pith:KUD6XAMB

load-bearing objection Useful dynamic GS simulation pipeline for navigation, but the central claim that photorealism drives the gains is not supported because the comparison with Vid2Sim is confounded. the 4 major comments →

arxiv 2602.11575 v3 pith:KUD6XAMB submitted 2026-02-12 cs.RO cs.AIcs.CV

ReaDy-Go: Real-to-Sim Dynamic 3D Gaussian Splatting Simulation for Environment-Specific Visual Navigation with Moving Obstacles

Seungyeon Yoo , Youngseok Jang , Dabin Kim , Youngsoo Han , Seungwoo Jung , H. Jin Kim This is my paper
classification cs.RO cs.AIcs.CV
keywords visual navigation3D Gaussian Splattingreal-to-sim simulationdynamic obstacleshuman avatar animationimitation learningsim-to-real transferenvironment-specific policy
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.

ReaDy-Go is a real-to-sim pipeline that builds a photorealistic simulation of a target deployment environment from one monocular video, then populates it with animated, photorealistic human avatars and uses it to train an RGB-only goal-navigation policy by imitation learning. The paper's central claim is that training in this dynamic, environment-specific simulator lets the policy transfer to the real world with substantially smaller performance drops than policies trained with non-photorealistic obstacle assets or general-purpose navigation models. In real-world dynamic tests across three environments, the ReaDy-Go policy reaches success rates of 90, 70, and 80 percent, and a zero-shot deployment in a fourth environment reaches 50 percent dynamic success. If correct, this makes environment-specific navigation policy training practical from just a few minutes of video, without depth sensors, mesh extraction, or a physics engine.

Core claim

The paper's central claim is that the sim-to-real gap in RGB-only navigation with moving obstacles can be largely closed by reconstructing the deployment scene as 3D Gaussian Splatting, inserting animatable human Gaussian avatars, and generating expert demonstrations from a planner that treats those avatars as dynamic obstacles. A static background is rebuilt from a short monocular video; human avatars extracted from separate videos are animated along 2D trajectories using a motion-generation prior; a search-based expert planner replans around projected, inflated human occupancy; and a compact convolutional policy is trained by imitation learning. The paper reports that this policy outperfor

What carries the argument

The load-bearing mechanism is the dynamic 3D Gaussian Splatting simulator: a static GS scene reconstructed with geometric regularization is combined with animatable human GS avatars and a motion module that converts 2D trajectories into plausible body motion. Around this simulator sit two planners. The robot expert planner voxelizes the static scene with opacity-based noise filtering, projects cleaned human Gaussian primitives onto a 2D navigable map, inflates them with a constant-velocity safety buffer, and replans with a motion-primitive search when the path becomes unsafe. The human planner samples crossing or parallel trajectories on a walkable map so training episodes are interaction-ri

Load-bearing premise

The load-bearing premise is that the metric-scale reconstruction of the deployment environment from a monocular video yields a 2D occupancy map accurate enough for the expert planner's paths and obstacle placements to match the real world; if the scale or geometry is off, the policy trains on a map of a different environment than the one it is deployed in.

What would settle it

Run a controlled comparison in one environment: train policy A with the full ReaDy-Go dynamic GS simulator and policy B with non-photorealistic mesh humans animated along identical trajectories in the same static GS scene, then deploy both on the same real dynamic test suite. The paper's mechanism predicts a clear success-rate gap in favor of A; if B matches A, the photorealism of human obstacles is not what carries the transfer. Alternatively, deliberately corrupt the metric scale of the reconstructed occupancy map by 20% during data generation and check whether real-world dynamic success col

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

If this is right

  • Environment-specific RGB-only navigation policies can be trained from a single monocular video of the deployment site, removing the need for manual digital-twin construction or physics-engine integration.
  • Photorealism of dynamic obstacles is a first-order factor in sim-to-real transfer for dynamic scenes: replacing them with mesh-based human assets measurably lowers dynamic success rates in the paper's comparisons.
  • A policy trained on several reconstructed environments can be deployed zero-shot in a new environment and retain meaningful navigation success, suggesting that scaling up the pipeline could produce general navigation models.
  • Because the policy needs only RGB frames and relative goal position, the same training pipeline can transfer across robot embodiments with a forward-facing camera and available odometry.
  • Expert data quality matters: reactive replanning against projected human Gaussian occupancy is what makes safe demonstrations available for imitation learning in dynamic GS scenes.

Where Pith is reading between the lines

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

  • The same simulator structure should extend to any articulated object that can be captured as an animatable Gaussian avatar (carts, animals, other robots), so the method is probably not limited to pedestrian-occupied scenes.
  • A likely failure mode the paper does not stress-test: the reconstruction's metric scale and occupancy accuracy come from monocular structure-from-motion plus a single marker; in scenes with reflective floors or glass walls, the expert planner's map could be wrong even though the rendered images look perfect, and the policy would inherit that error.
  • Testable extension: adding domain randomization over lighting, floor texture, and camera noise inside the GS simulator could close more of the remaining dynamic-task gap without sacrificing photorealism.
  • The zero-shot result is a hint rather than a guarantee: 50% dynamic success in an unseen environment suggests the policy learns general obstacle-detouring behavior, but reaching practical reliability would likely require many more training environments.

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. The paper introduces ReaDy-Go, a real-to-sim pipeline for training RGB-only visual navigation policies in dynamic environments. A static 3D Gaussian Splatting (GS) scene is reconstructed from a monocular video, dynamic human obstacles are inserted as animatable GS avatars driven by a motion generation module, and a Hybrid A* expert planner plus a human path planner are used to generate photorealistic navigation datasets. A lightweight CNN+MLP policy is trained by imitation learning on these datasets. The authors evaluate ReaDy-Go in three target environments in simulation (100 episodes per condition) and on a real differential-drive robot (10 episodes per condition), comparing against Vid2Sim, GNM, ViNT, and NoMaD. They report higher success rates and lower average reaching times than baselines, especially in dynamic tasks, and present a zero-shot deployment in an unseen environment that achieves roughly 50% success. The central claim is that photorealistic dynamic GS obstacles are the key factor behind the improved dynamic-task performance.

Significance. If the central claim holds, ReaDy-Go is a useful step toward environment-specific visual navigation in dynamic scenes: it avoids mesh extraction and physics-engine integration, generates photorealistic dynamic data from a single video, and shows that GS-based real-to-sim training can transfer to a real robot. The paper's strengths are its concrete pipeline (static GS + animatable human GS + planners), a plausible data-generation mechanism, and real-robot validation with multiple baselines. However, the main causal attribution — that photorealism of the dynamic obstacles drives the improvement over Vid2Sim — is not yet supported because the comparison is confounded by differences in trajectory generation, obstacle dynamics, and the expert planner. The real-world evidence is also statistically fragile (n=10 per condition). The contribution is contingent on resolving these issues.

major comments (4)
  1. [§IV.D, §III-B.2–3, Table I] The paper claims (Section IV.D) that ReaDy-Go and Vid2Sim 'differ only in the training data' and attributes ReaDy-Go's higher dynamic success rate to photorealistic human GS obstacles. This comparison is confounded: the two pipelines differ in (i) the human trajectory generation (ReaDy-Go uses A* + spline optimization and PriorMDM; Vid2Sim uses Unity physics-engine assets), (ii) the robot expert planner (ReaDy-Go uses a bespoke Hybrid A* planner with reactive replanning), and (iii) obstacle dynamics, collision geometry, speeds, and possibly obstacle counts. Any of these differences, not photorealism, could explain the observed dynamic-task gap. The manuscript needs an experiment that holds obstacle trajectories and robot start/goal pairs fixed across simulators, or an ablation within ReaDy-Go that swaps GS humans for non-photorealistic meshes while keeping the rest of the pipeline fixed.
  2. [§IV.E, Table II] Real-world results are based on 10 episodes per condition. The reported differences, e.g., 90% vs 60% dynamic success in Outside, correspond to 3 episodes out of 10 and are not statistically reliable; no confidence intervals, significance tests, or multiple training seeds are reported. Simpson's-paradox-style conclusions from such small samples are risky. The authors should provide exact binomial confidence intervals or run more episodes, and should also report variance over policy training seeds in the simulation experiments. Without this, the real-world claims are not firmly supported.
  3. [§IV.F, Table III] The zero-shot deployment result is presented as evidence of generalization potential, but the unseen-environment success rates (70% static, 50% dynamic) are based on a small number of episodes and are not compared to any baseline or chance level. A 50% dynamic success rate in an unseen environment, without confidence intervals or a comparison to a randomly initialized policy, is too weak to support the claim in the abstract that the method shows 'generalization potential.' Please provide the number of episodes, confidence intervals, and ideally a comparison to the same policy without environment-specific training.
  4. [§III-B.1, §III-B.2] Several pipeline thresholds and design choices are described qualitatively but without values or sensitivity analysis: the opacity threshold for free-space classification, the safety margin for inflation, the 75th-percentile covariance-trace threshold, the 2 s constant-velocity horizon, and the replanning trigger. Since the pipeline's effectiveness depends on these parameters, the paper should report the chosen values and at least a brief sensitivity study. Without this, the reproducibility of the method is limited and it is unclear how robust the reported gains are to these choices.
minor comments (6)
  1. [Table I caption] The 'Params' column lists model sizes but for the Expert row it is empty; please clarify that the expert is not a learned policy. Also, the table would benefit from noting the number of episodes per simulation condition (stated in the text as 100).
  2. [§III-B.1] The phrase '1σ Gaussian ellipsoid' is not defined. Please specify the covariance scaling used when deciding voxel occupancy.
  3. [§IV.B.1] The statement 'we generated 400 training episodes for each environment' is clear, but the validation episodes (50) are used for checkpoint selection; it would be helpful to state whether the reported simulation results are from the best validation checkpoint, and how many training seeds were used.
  4. [§IV.D] The sentence 'Since the two methods differ only in the training data...' is overclaimed, as detailed in Major Comment 1. Please soften to 'In our comparison, the two methods were trained with the same policy architecture but different data-generation pipelines.'
  5. [§IV.F] The phrase 'over a 50% success rate' is imprecise given Table III reports exactly 50% for the dynamic task. Please report the exact number and its uncertainty.
  6. [General] No code or trained model is released, which limits reproducibility. The project page is mentioned, but please clarify in the paper what artifacts (code, datasets, checkpoints) will be made available.

Circularity Check

0 steps flagged

No significant circularity: ReaDy-Go is an empirical pipeline evaluation whose claims are not derived from fitted inputs or self-citations.

full rationale

This paper is an empirical systems paper rather than a derivation, and I found no circular step that reduces a stated result to its own inputs. The expert planner, human planner, voxelization thresholds, and animation modules are heuristic pipeline components; none of their parameters are fitted to make the reported navigation metrics come out favorably. The training data are generated by the pipeline and the policy is evaluated on held-out episodes, so the success-rate comparisons are external to the data-generation procedure. The only self-citation by the authors, [19], appears in related work on cross-modal learning and is not load-bearing for the central claims. The strongest validity concern is the paper's assertion that ReaDy-Go and Vid2Sim 'differ only in the training data' when they in fact differ in obstacle representation, trajectory generation, and planner details; this is a confounding-variable problem for the causal attribution to photorealism, not circularity. The key results, including real-world success rates, are empirical outcomes that could in principle fail, and they are not guaranteed by construction. Therefore the appropriate circularity score is 0.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The pipeline depends on a set of hand-chosen thresholds and heuristic parameters (opacity threshold, safety margins, prediction horizon, percentile cutoffs) that are not systematically tuned or justified by ablations. The method also relies on several pre-existing learned models (PGSR, HUGS, PriorMDM, SMPLify) as black-box components, each carrying its own assumptions about accuracy and transferability. No new physical entities, forces, or conserved quantities are introduced. The central claim is an empirical one, supported by the experimental setup but limited by the small real-world test set and missing implementation details.

free parameters (7)
  • Opacity threshold for voxel free-space classification = Not reported
    Used in Section III-B.1 to filter spurious Gaussians during voxelization; affects the robot's and human's navigable maps and thus the quality of expert trajectories.
  • Safety margin for robot/human inflation = Not reported
    Inflates occupied grids by robot and human radii in Section III-B.1; directly affects the feasibility and conservativeness of planned paths.
  • Constant-velocity prediction horizon = 2 s
    Section III-B.2 uses a 2 s constant-velocity model to inflate human-occupied regions; a key parameter in the dynamic collision-checking.
  • 75th percentile covariance trace threshold = 75th percentile of per-human covariance trace
    Used in Section III-B.2 to discard spurious human Gaussian primitives; the choice of percentile is heuristic and not justified by an ablation.
  • Planner cost weights and replan period = Not reported
    The Hybrid A* cost is a weighted sum of arc length, steering penalty, and heuristic cost-to-go; the weights and the periodic replanning interval are not reported, affecting the optimality and safety of expert demonstrations.
  • Motion primitive velocity scales and steering commands = 1/3, 2/3, 1 of max velocity; left, straight, right
    Discrete motion primitive library for Hybrid A* (Section III-B.2); the discretization can influence whether the expert planner finds a feasible path in tight dynamic scenarios.
  • Training hyperparameters of the policy = Learning rate 1e-4, image resolution 144x256, action interval 0.5s; batch size not reported
    The policy architecture and hyperparameters (Section IV-B.2) are chosen without systematic tuning; batch size and training length are not reported, impacting reproducibility and final performance.
axioms (6)
  • domain assumption PGSR reconstructs an accurate metric-scale static scene from a monocular video with COLMAP and an ArUco marker.
    Stated in Section III-A.1. The geometric fidelity of the reconstructed scene underpins the voxelization, occupancy maps, expert planning, and the final sim-to-real transfer.
  • domain assumption HUGS-extracted human GS avatars can be re-animated under novel poses and rendered in novel scenes with sufficient fidelity.
    Section III-A.2 assumes the human avatars from the NeuMan dataset, fitted with HUGS, transfer to arbitrary scenes and viewpoints without degradation.
  • domain assumption PriorMDM-generated human motions, after SMPLify fitting, are plausible and realistic enough to train a navigation policy that generalizes to real humans.
    The motion generation module (Section III-A.2) relies on PriorMDM's generative prior and SMPLify fitting; the paper does not quantify the visual or kinematic realism of the resulting animations.
  • domain assumption The occupied-voxel map derived from Gaussian ellipsoids with opacity filtering is an accurate enough representation of the environment for safe planning.
    Section III-B.1 describes voxelization and opacity filtering; if weak-textured regions or spurious Gaussians cause incorrect free/occupied classifications, both robot and human plans will be flawed.
  • domain assumption The differential-drive robot's motion is well approximated by a unicycle model with the motion primitives used in Hybrid A*.
    Section IV-A.1 states the robot is a differential wheeled robot; the expert planner's primitives (Section III-B.2) are discrete and the paper does not verify that the robot's actual kinematics lie within the primitive set.
  • domain assumption A dataset of 400 training episodes per environment is sufficient to cover the distribution of static and dynamic scenarios the policy will encounter.
    Section IV-B.1 reports 400 episodes per environment; there is no analysis of dataset size or coverage, so the policy may overfit to the sampled start/goal/human-trajectory distribution.

pith-pipeline@v1.3.0-alltime-deepseek · 11863 in / 10756 out tokens · 157213 ms · 2026-08-03T00:03:50.251262+00:00 · methodology

0 comments
read the original abstract

Visual navigation models often struggle in real-world dynamic environments due to limited robustness to the sim-to-real gap and the difficulty of training policies tailored to target deployment environments (e.g., households, restaurants, and factories). Although real-to-sim navigation simulation using 3D Gaussian Splatting (GS) can mitigate these challenges, prior GS-based works have considered only static scenes or non-photorealistic human obstacles built from simulator assets, despite the importance of safe navigation in dynamic environments. To address these issues, we propose ReaDy-Go, a novel real-to-sim simulation pipeline that synthesizes photorealistic dynamic scenarios in target environments by augmenting a reconstructed static GS scene with dynamic human GS obstacles, and trains navigation policies using the generated datasets. The pipeline provides three key contributions: (1) a dynamic GS simulator that integrates static scene GS with a human animation module, enabling the insertion of animatable human GS avatars and the synthesis of plausible human motions from 2D trajectories, (2) a navigation dataset generation framework that leverages the simulator along with a robot expert planner designed for dynamic GS representations and a human planner, and (3) robust navigation policies to both the sim-to-real gap and moving obstacles. The proposed simulator generates thousands of photorealistic navigation scenarios with animatable human GS avatars from arbitrary viewpoints. ReaDy-Go outperforms baselines across target environments in both simulation and real-world experiments, demonstrating improved navigation performance even after sim-to-real transfer and in the presence of moving obstacles. Moreover, zero-shot sim-to-real deployment in an unseen environment indicates its generalization potential. Project page: https://syeon-yoo.github.io/ready-go-site/.

Figures

Figures reproduced from arXiv: 2602.11575 by Dabin Kim, H. Jin Kim, Seungwoo Jung, Seungyeon Yoo, Youngseok Jang, Youngsoo Han.

Figure 1
Figure 1. Figure 1: The proposed real-to-sim dynamic environment sim￾ulation pipeline for visual navigation. ReaDy-Go generates photorealistic navigation datasets for dynamic scenarios and trains environment-specific visual navigation policies from these datasets. The resulting policies demonstrate robustness to the sim-to-real gap and moving obstacles. extensively in simulation, as collecting large-scale real-world navigatio… view at source ↗
Figure 2
Figure 2. Figure 2: ReaDy-Go overview. The proposed photorealistic simulation pipeline for visual navigation in dynamic environments consists of three main components: (1) a real-to-sim dynamic 3D Gaussian Splatting (GS) simulator with animatable human GS avatars, (2) photorealistic navigation dataset generation for dynamic scenarios, and (3) visual navigation policy training. In contrast, our method, ReaDy-Go, develops a pip… view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of the robot expert planner. (a) The robot follows a collision-free path (red) from start (green) to goal (blue). (b) When a dynamic obstacle (human point cloud in red; inflated region in magenta) makes the path unsafe, the robot follows a replanned path (yellow). better corresponds to the actual human-occupied region for 2D projection, we remove spurious primitives by discarding ellipsoids w… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative novel-view synthesis results from the proposed dynamic GS simulation pipeline across diverse viewpoints and environments. ReaDy-Go generates photorealistic, geometrically consistent dynamic scenarios with natural human motion from novel viewpoints, enabling navigation dataset generation for target deployment environments. TABLE I: Visual Navigation Performance in Simulation Method Params Outsid… view at source ↗
Figure 5
Figure 5. Figure 5: Real-world experiments. ReaDy-Go demonstrates robust real-world visual navigation performance after sim￾to-real transfer in target environments (a–c) and an unseen environment (d), in both Static and Dynamic. Second, photorealistic dynamic obstacles in our simu￾lation are a key factor in maintaining visual navigation performance in dynamic environments. While ReaDy-Go and Vid2Sim perform comparably in stat… view at source ↗

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. Image2Sim: Scaling Embodied Navigation via Generative Neural Simulator

    cs.CV 2026-07 conditional novelty 6.5

    A feed-forward feature-Gaussian plus one-step geometry-aware pixel-flow simulator converts large image collections into 20K interactive scenes and 10M+ navigation samples that improve zero-shot Habitat and real-robot ...

  2. DeGS: A Scalable 3DGS Architecture via Decoupled Workload Parsing and Reorganization

    cs.AR 2026-08 conditional novelty 6.0

    DeGS restructures 3DGS rendering into span parsing, task reorganization, and dense blending stages, achieving 1.8x-7.2x speedup and >80% scaling utilization over prior 3DGS accelerators.

Reference graph

Works this paper leans on

37 extracted references · 5 linked inside Pith · cited by 2 Pith papers

  1. [1]

    GNM: A General Navigation Model to Drive Any Robot,

    D. Shah, A. Sridhar, A. Bhorkar, N. Hirose, and S. Levine, “GNM: A General Navigation Model to Drive Any Robot,” inInternational Conference on Robotics and Automation (ICRA), 2023

    2023

  2. [2]

    ViNT: A foundation model for visual navigation,

    D. Shah, A. Sridhar, N. Dashora, K. Stachowicz, K. Black, N. Hirose, and S. Levine, “ViNT: A foundation model for visual navigation,” in 7th Annual Conference on Robot Learning, 2023

    2023

  3. [3]

    Nomad: Goal masked diffusion policies for navigation and exploration,

    A. Sridhar, D. Shah, C. Glossop, and S. Levine, “Nomad: Goal masked diffusion policies for navigation and exploration,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 63–70

    2024

  4. [4]

    Embodiedsplat: Personalized real-to-sim-to-real navigation with gaussian splats from a mobile device,

    G. Chhablani, X. Ye, M. Z. Irshad, and Z. Kira, “Embodiedsplat: Personalized real-to-sim-to-real navigation with gaussian splats from a mobile device,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2025, pp. 25 431– 25 441

    2025

  5. [5]

    Where are we in the search for an artificial visual cortex for embodied intelligence?

    A. Majumdar, K. Yadav, S. Arnaud, J. Ma, C. Chen, S. Silwal, A. Jain, V .-P. Berges, T. Wu, J. Vakil, P. Abbeel, J. Malik, D. Batra, Y . Lin, O. Maksymets, A. Rajeswaran, and F. Meier, “Where are we in the search for an artificial visual cortex for embodied intelligence?” inAd- vances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globers...

    2023

  6. [6]

    Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI,

    S. K. Ramakrishnan, A. Gokaslan, E. Wijmans, O. Maksymets, A. Clegg, J. M. Turner, E. Undersander, W. Galuba, A. Westbury, A. X. Chang, M. Savva, Y . Zhao, and D. Batra, “Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI,” inThirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021

    2021

  7. [7]

    Vr- robo: A real-to-sim-to-real framework for visual robot navigation and locomotion,

    S. Zhu, L. Mou, D. Li, B. Ye, R. Huang, and H. Zhao, “Vr- robo: A real-to-sim-to-real framework for visual robot navigation and locomotion,”IEEE Robotics and Automation Letters, vol. 10, no. 8, pp. 7875–7882, 2025

    2025

  8. [8]

    Gaussian splatting to real world flight navigation transfer with liquid networks,

    A. Quach, M. Chahine, A. Amini, R. Hasani, and D. Rus, “Gaussian splatting to real world flight navigation transfer with liquid networks,” in8th Annual Conference on Robot Learning, 2024

    2024

  9. [9]

    Sous vide: Cooking visual drone navigation policies in a gaussian splatting vacuum,

    J. Low, M. Adang, J. Yu, K. Nagami, and M. Schwager, “Sous vide: Cooking visual drone navigation policies in a gaussian splatting vacuum,”IEEE Robotics and Automation Letters, vol. 10, no. 5, pp. 5122–5129, 2025

    2025

  10. [10]

    Performance-guided refinement for visual aerial navigation using editable gaussian splatting in falcongym 2.0,

    Y . Miao, E. Yuceel, G. Fainekos, B. Hoxha, H. Okamoto, and S. Mitra, “Performance-guided refinement for visual aerial navigation using editable gaussian splatting in falcongym 2.0,”arXiv preprint arXiv:2510.02248, 2025

    arXiv 2025

  11. [11]

    Vid2sim: Realistic and interactive simulation from video for urban navigation,

    Z. Xie, Z. Liu, Z. Peng, W. Wu, and B. Zhou, “Vid2sim: Realistic and interactive simulation from video for urban navigation,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2025, pp. 1581–1591

    2025

  12. [12]

    Sketchplan: Dif- fusion based drone planning from human sketches,

    S. Norelius, A. O. Feldman, and M. Schwager, “Sketchplan: Dif- fusion based drone planning from human sketches,”arXiv preprint arXiv:2510.03545, 2025

    arXiv 2025

  13. [13]

    3d gaussian splatting for real-time radiance field rendering,

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering,”ACM Transactions on Graphics, vol. 42, no. 4, July 2023

    2023

  14. [14]

    Visual navigation among humans with optimal control as a supervisor,

    V . Tolani, S. Bansal, A. Faust, and C. Tomlin, “Visual navigation among humans with optimal control as a supervisor,”IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 2288–2295, 2021

    2021

  15. [15]

    Deep drone racing: From simulation to reality with domain randomization,

    A. Loquercio, E. Kaufmann, R. Ranftl, A. Dosovitskiy, V . Koltun, and D. Scaramuzza, “Deep drone racing: From simulation to reality with domain randomization,”IEEE Transactions on Robotics, vol. 36, no. 1, pp. 1–14, 2020

    2020

  16. [16]

    Cad2rl: Real single-image flight without a single real image,

    F. Sadeghi and S. Levine, “Cad2rl: Real single-image flight without a single real image,”arXiv preprint arXiv:1611.04201, 2016

    Pith/arXiv arXiv 2016

  17. [17]

    Offline visual representation learning for embodied navigation,

    K. Yadav, R. Ramrakhya, A. Majumdar, V .-P. Berges, S. Kuhar, D. Ba- tra, A. Baevski, and O. Maksymets, “Offline visual representation learning for embodied navigation,” inWorkshop on Reincarnating Reinforcement Learning at ICLR 2023, 2023

    2023

  18. [18]

    Ovrl-v2: A simple state-of-art baseline for imagenav and objectnav,

    K. Yadav, A. Majumdar, R. Ramrakhya, N. Yokoyama, A. Baevski, Z. Kira, O. Maksymets, and D. Batra, “Ovrl-v2: A simple state-of-art baseline for imagenav and objectnav,”arXiv preprint arXiv:2303.07798, 2023

    Pith/arXiv arXiv 2023

  19. [19]

    Mono-camera-only target chasing for a drone in a dense environment by cross-modal learning,

    S. Yoo, S. Jung, Y . Lee, D. Shim, and H. J. Kim, “Mono-camera-only target chasing for a drone in a dense environment by cross-modal learning,”IEEE Robotics and Automation Letters, vol. 9, no. 8, pp. 7254–7261, 2024

    2024

  20. [20]

    Learn- ing visuomotor policies for aerial navigation using cross-modal repre- sentations,

    R. Bonatti, R. Madaan, V . Vineet, S. Scherer, and A. Kapoor, “Learn- ing visuomotor policies for aerial navigation using cross-modal repre- sentations,” in2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020, pp. 1637–1644

    2020

  21. [21]

    Urbanverse: Scal- ing urban simulation by watching city-tour videos,

    M. Liu, H. He, E. Ricci, W. Wu, and B. Zhou, “Urbanverse: Scal- ing urban simulation by watching city-tour videos,”arXiv preprint arXiv:2510.15018, 2025

    arXiv 2025

  22. [22]

    Phone2proc: Bringing robust robots into our chaotic world,

    M. Deitke, R. Hendrix, A. Farhadi, K. Ehsani, and A. Kembhavi, “Phone2proc: Bringing robust robots into our chaotic world,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 9665–9675

    2023

  23. [23]

    Splatsim: Zero-shot sim2real transfer of rgb manipulation policies using gaussian splatting,

    M. N. Qureshi, S. Garg, F. Yandun, D. Held, G. Kantor, and A. Silwal, “Splatsim: Zero-shot sim2real transfer of rgb manipulation policies using gaussian splatting,”arXiv preprint arXiv:2409.10161, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Real2render2real: Scaling robot data without dynamics simulation or robot hardware,

    J. Yu, L. Fu, H. Huang, K. El-Refai, R. A. Ambrus, R. Cheng, M. Z. Irshad, and K. Goldberg, “Real2render2real: Scaling robot data without dynamics simulation or robot hardware,”arXiv preprint arXiv:2505.09601, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    Robo-gs: A physics consistent spatial-temporal model for robotic arm with hybrid representation,

    H. Lou, Y . Liu, Y . Pan, Y . Geng, J. Chen, W. Ma, C. Li, L. Wang, H. Feng, L. Shi, L. Luo, and Y . Shi, “Robo-gs: A physics consistent spatial-temporal model for robotic arm with hybrid representation,” in2025 IEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 15 379–15 386

    2025

  26. [26]

    Splat-nav: Safe real-time robot navi- gation in gaussian splatting maps,

    T. Chen, O. Shorinwa, J. Bruno, A. Swann, J. Yu, W. Zeng, K. Nagami, P. Dames, and M. Schwager, “Splat-nav: Safe real-time robot navi- gation in gaussian splatting maps,”IEEE Transactions on Robotics, vol. 41, pp. 2765–2784, 2025

    2025

  27. [27]

    Gaussnav: Gaussian splatting for visual navigation,

    X. Lei, M. Wang, W. Zhou, and H. Li, “Gaussnav: Gaussian splatting for visual navigation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 5, pp. 4108–4121, 2025

    2025

  28. [28]

    Structure-from-motion revisited,

    J. L. Sch ¨onberger and J.-M. Frahm, “Structure-from-motion revisited,” in2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 4104–4113

    2016

  29. [29]

    Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction,

    D. Chen, H. Li, W. Ye, Y . Wang, W. Xie, S. Zhai, N. Wang, H. Liu, H. Bao, and G. Zhang, “Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction,”IEEE Transactions on Visualization and Computer Graphics, vol. 31, no. 9, pp. 6100– 6111, 2025

    2025

  30. [30]

    Neuman: Neural human radiance field from a single video,

    W. Jiang, K. M. Yi, G. Samei, O. Tuzel, and A. Ranjan, “Neuman: Neural human radiance field from a single video,” inProceedings of the European conference on computer vision (ECCV), 2022

    2022

  31. [31]

    HUGS: Human gaussian splatting,

    M. Kocabas, J.-H. R. Chang, J. Gabriel, O. Tuzel, and A. Ranjan, “HUGS: Human gaussian splatting,” in2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. [Online]. Available: https://arxiv.org/abs/2311.17910

    Pith/arXiv arXiv 2024

  32. [32]

    SMPL: A skinned multi-person linear model,

    M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black, “SMPL: A skinned multi-person linear model,”ACM Trans. Graphics (Proc. SIGGRAPH Asia), vol. 34, no. 6, pp. 248:1–248:16, Oct. 2015

    2015

  33. [33]

    Human motion diffusion as a generative prior,

    Y . Shafir, G. Tevet, R. Kapon, and A. H. Bermano, “Human motion diffusion as a generative prior,” inThe Twelfth International Confer- ence on Learning Representations, 2024

    2024

  34. [34]

    Generating diverse and natural 3d human motions from text,

    C. Guo, S. Zou, X. Zuo, S. Wang, W. Ji, X. Li, and L. Cheng, “Generating diverse and natural 3d human motions from text,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022, pp. 5152–5161

    2022

  35. [35]

    Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image,

    F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero, and M. J. Black, “Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image,” inComputer Vision – ECCV 2016, ser. Lecture Notes in Computer Science. Springer International Publishing, Oct. 2016

    2016

  36. [36]

    Practical search techniques in path planning for autonomous driving,

    D. Dolgov, S. Thrun, M. Montemerlo, and J. Diebel, “Practical search techniques in path planning for autonomous driving,”ann arbor, vol. 1001, no. 48105, pp. 18–80, 2008

    2008

  37. [37]

    A history of the unity game engine,

    J. K. Haas, “A history of the unity game engine,” 2014

    2014