Pith. sign in

REVIEW 5 major objections 5 minor 37 references

EmbodiedDiffusion: End-to-End Traversability-Guided Visual Diffusion for Heterogeneous Robot Navigation

T0 review · 5 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read One diffusion model outputs both a traversability map and a feasible trajectory from a single RGB image, transferring across a quadruped and a drone with 80–100% success.

desk verdict Plausible and genuinely integrated system, but the headline claims outrun the evidence: unablated APF/post-hoc corrections and teacher-in-the-loop metrics mean the end-to-end planner-free story is not yet demonstrated. read the letter →

arxiv 2512.02851 v4 pith:NZFD63SL submitted 2025-12-02 cs.RO

Iana Zhura , Sausar Karaf , Faryal Batool , Nipun Dhananjaya Weerakkodi Mudalige , Valerii Serpiva , Ali Alridha Abdulkarim , Aleksey Fedoseev , Didar Seyidov
show 2 more authors
Hajira Amjad Dzmitry Tsetserukou
This is my paper
classification cs.RO
keywords traversabilityestimationdiffusionmodeltrajectorygenerationcross-embodimentnavigationRGB-onlyplanner-freetrainingembodimentconditioningreal-timerobotinference
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that visual navigation can be collapsed into a single generative model that reads one RGB image and directly outputs both a traversability map and a drivable trajectory, with no prompts, no expert demonstrations, and no external planner. If right, robots of different shapes and dynamics could be reprogrammed for navigation with only a few hundred new images, because the embodiment-specific knowledge is isolated in a small trainable conditioning module. The authors support the claim with indoor experiments on a legged robot and a custom drone, reporting real-time inference at 0.09 seconds and successful deployment in unseen environments. The core idea is that randomized, synthetic path skeletons plus traversability-aware regularization are enough to teach a diffusion model stable motion priors, eliminating the need for planner-generated supervision.

What carries the argument

The central object is a conditional diffusion UNet that operates on 64×64 trajectory heatmaps, conditioned on three aligned spatial signals: the noisy trajectory, a start–goal tensor, and the predicted traversability map, plus FiLM-modulated visual tokens carrying embodiment information. The traversability student model is a frozen visual backbone with a trainable adapter and a FiLM layer that scales and shifts image features by a six-dimensional robot state, so embodiment-specific reasoning is isolated from the shared visual encoder. The trajectory supervision is generated by a planner-free pipeline: random waypoints, Bézier smoothing, thin rasterization, and a set of losses (directional fo

What would settle it

Run the full deployment pipeline on the quadruped and drone in the same indoor test environments but disable the APF refinement and the cone-shaped visibility mask entirely, using only the raw diffusion-model trajectory projected to 3D. If navigation success drops substantially (e.g., from 80–100% to near chance) or collisions occur that the APF was previously preventing, then the end-to-end planner-free claim fails because the safety layer is doing the load-bearing planning work.

Watch

Extended reading notes

Core claim

The paper's central claim is that a lightweight end-to-end diffusion model can jointly predict a traversability map and generate a feasible trajectory from a single RGB image, using a 2D image-space representation and a compact embodiment state vector. The model is trained without expert trajectories: synthetic supervision comes from Gaussian-sampled waypoints interpolated with Bézier curves and rasterized into thin heatmaps, regularized by directional, traversability, and mass-based losses. A VLM teacher provides traversability labels during training only, so inference is prompt-free. A FiLM-based conditioning mechanism modifies visual features according to the robot's state vector, allowin

Load-bearing premise

The load-bearing assumption is that the artificial potential field refinement and any post-hoc traversability-map modifications are only minor safety corrections, not the actual reason the robot avoids collisions; the paper does not test the diffusion model's trajectories without these layers.

Editorial extensions

If this is right

  • If the claim holds, robot navigation could be treated as a single-image conditional generation problem, removing the need for modular pipelines that build costmaps and run separate planners.
  • The same pretrained visual backbone and trajectory diffusion model could be reused across robot platforms, with adaptation reduced to learning a small FiLM conditioning head from a few hundred images.
  • Prompt-free inference means deployment no longer depends on tuning vision-language prompts at runtime, which the paper identifies as a major fragility of existing VLM-based traversability methods.
  • The reported 0.09 s inference at 20 denoising steps suggests the architecture is light enough for real-time onboard use on modest GPU hardware, not just off-board workstations.
  • The observation that small post-hoc modifications to the traversability map (like a cone-shaped visibility mask) steer the trajectory implies the model is controllable through its conditioning input, which could simplify safe operation in constrained spaces.

Reading between the lines

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

  • A reader might infer that the model is not truly planning from pure vision but learning a strong visual prior that the APF safety layer then corrects; the paper does not ablate away that layer, so the end-to-end claim is only as strong as the assumption that the safety layer is a minor correction.
  • The reported success rates are heavily concentrated in structured indoor scenarios (gates, corridors); it is an open question whether the same single-image approach would transfer to outdoor, unstructured, or highly dynamic environments where navigation needs more than a static traversability prior.
  • The steerability through traversability-map edits suggests a practical interface for human or geometric constraints at inference time, but it also means the model's outputs are sensitive to the quality of the predicted map—errors in the traversability prediction could directly corrupt the trajectory without an independent planner to catch them.
  • If the method scales to long-horizon and multi-step planning, the same image-conditioned diffusion framework could potentially unify local and global navigation, but the current experiments are limited to short indoor routes and do not yet demonstrate memory or sequential decision-making.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes SwarmDiffusion (submission title EmbodiedDiffusion), a diffusion-based framework that jointly predicts traversability maps and generates feasible trajectories from a single RGB image. It distills traversability from a VLM teacher (AnyTraverse) into a lightweight student, uses FiLM conditioning for embodiment, and trains the trajectory generator without expert demonstrations, using randomized waypoint sampling with Bézier smoothing as synthetic supervision. Experiments are reported on a Unitree Go1 quadruped and a custom quadrotor, with 80–100% success in full-data regimes, 0.09 s inference, and simulation results in Unity. The paper claims a planner-free, prompt-free, annotation-free approach with cross-embodiment transfer.

Significance. If the central claims hold, the work would be a meaningful step toward unified traversability reasoning and trajectory generation, removing the need for separate planners and prompt engineering, and enabling quick adaptation to new robot platforms. The real-hardware experiments are a strength, as is the attempt to distill VLM semantics into a small, deployable model. However, the evidence is weakened by missing ablations of the safety layer, small numbers of trials, and an evaluation that partly uses the same teacher that generated the supervision. The idea is promising, but the current manuscript does not yet substantiate the planner-free and end-to-end claims.

major comments (5)
  1. [§V.J and §I] The central 'planner-free' claim is not established. Real-world results (Tables VII and VIII) are reported with a repulsive APF force applied to all trajectory points (Eq. 35) and with a post-hoc cone-shaped visibility mask on the traversability map (§V.J.1). No ablation removes these corrections and evaluates the raw diffusion output. If the APF or mask is the primary source of collision avoidance, the 'end-to-end planner-free' claim fails. The authors should ablate both corrections and report success, clearance, and safety metrics for the raw model output.
  2. [§V.F, Eq. (24)] The simulation and ablation metrics use the teacher-derived traversability map T and cost C(x) as the reference, while the same teacher supplies the supervision signal for L_distil (Eq. 24) and is used to select goals in dataset generation (§V.D). Thus, the reported cost and safety improvements partly measure agreement with the very model that generated the training signal. This circularity affects Tables I–VI. The authors should use an independent ground truth (e.g., point-cloud collision checks) or clearly discuss the limitation.
  3. [§IV.B.4, Eq. (23)] The mass-based regularization L_mass is written as (1/HW) Σ p(u,v). Since p is defined via a global softmax (Eq. 20), its sum is exactly 1, making L_mass constant and ineffective. The text describes penalties for obstacle proximity, path thickness, entropy compactness, and spatial smoothness, but these are not represented in the equation. The training objective is therefore mis-specified, and the actual regularizers used in training are unclear. The correct equations must be provided.
  4. [§V.H, Table V] The baseline comparison is not fully fair. The baseline is evaluated at 50% data for the dog and 100% for the drone, while the paper claims superiority 'even at 30% data' without a baseline at that fraction. Also, the baseline uses A* expert trajectories, while the proposed method uses synthetic non-expert trajectories; this is a fundamental difference in supervision. The quantitative comparison in Table V reports MSE, which is by construction higher for the proposed method. Success rates or task-completion metrics beyond MSE should be reported for an apples-to-apples comparison.
  5. [§V.J, Tables VII and VIII] Real-robot experiments use only three trials per condition (or fewer), giving success rates of 67–100% with no error bars or statistical significance. With n=3, a single failure changes the success rate by 33 percentage points. The abstract's '80–100%' claim is based on very small samples, and the 67% results in the 30% dog and 70% drone regimes are not reflected in the summary. More trials and variance reporting are needed to support the success-rate claims.
minor comments (5)
  1. [Title] The submission title is 'EmbodiedDiffusion' but the manuscript text uses 'SwarmDiffusion' throughout. The inconsistency must be resolved.
  2. [Abstract and Introduction] The abstract claims adaptation using '10 min of visual data collection' and the introduction claims '500 additional visual samples,' but the experiments use 50% (2000 images) and 30% (1200 images) data regimes. The 500-sample claim is not directly tested.
  3. [Eq. (22)] The traversability expectation reward uses an undefined index 'b' (likely batch or sample index). Please clarify.
  4. [References] References [23] and [30] appear to be the same VENTURA paper; please deduplicate and check the citation list.
  5. [Fig. 2] Figure 2 is referenced but not fully described in the text; panels and conditioning pathways should be explained in the caption or body.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: traversability-based metrics are computed from the same VLM teacher used to supervise the student, but real-robot success provides independent support.

  1. self definitional [§V.F (Ablation Metrics, Eq. 33), §IV.B.5/7 (Eqs. 22, 24)]
    "L_trav = 1−max_b E_{p_b}[\hat T_b] ... L_distil = 1/HW \sum_{u,v} |\hat T(u,v)−T(u,v)| ... We compute all metrics directly from the generated trajectory τ and the predicted traversability map \hat T, using the ground-truth traversability map T as reference. ... C(x) the teacher-derived traversability cost."

    The student traversability map \hat T is trained with L_distil to match the VLM teacher T, and trajectory training rewards high E_p[\hat T] via L_trav. The reported Cost metric is evaluated with teacher-derived C(x), and the metric definition states that T is the reference. Thus the traversability-based simulation/ablation numbers partly measure agreement with the exact teacher that supplied the training signal and goal selection; lower cost is to some degree built in rather than independently demonstrated. Real-robot success and collision outcomes remain independent evidence.

full rationale

The core derivation is not circular in its training mechanism: randomized B-spline waypoints plus diffusion denoising are genuine synthetic supervision, and no claim is reduced to a self-citation. AnyTraverse is used as an external teacher, not an output of this paper. The main circularity concern is the evaluation loop: the teacher-created traversability map supervises the student (Eq. 24), conditions trajectory generation (Eqs. 11 and 22), and also defines the evaluation cost (Eq. 33). This makes the Cost and traversability-related metrics partially self-referential. The APF repulsion and post-hoc cone masks are unablated deployment corrections, but that is a correctness/attribution gap, not a circularity. The 80–100% real-robot success rates and collision-free execution are independent physical checks, so the paper retains substantial non-circular content. Score 4 rather than 0–2 because one central evaluation metric is tied to the training teacher by construction.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities; it assembles existing components (VLM teacher, DINOv2, FiLM, DDPM, APF) into a new architecture. The central claim rests primarily on unverified assumptions about the teacher's correctness, the adequacy of synthetic trajectories, and the smallness of the APF correction.

free parameters (5)
  • number of denoising steps = 20
    Chosen by ablation in §V.G.3; 20 steps maximize clearance/cost metrics on the validation split.
  • traversability threshold δ
    Used in the safety constraint Eq. (9b) but its value is never reported.
  • softmax temperature T
    Appears in the path probability distribution Eq. (20); no value is given.
  • APF separation coefficient c_s and sphere radius
    Parameters in Eq. (35) for the safety refinement layer; not reported.
  • loss weights for Ldir, Ltrav, Lmass, Ldistil
    The combined loss Eq. (26) is written without explicit weights, but such weights are almost certainly tuned; they are not reported.
assumptions (5)
  • domain assumption AnyTraverse teacher traversability maps are reliable enough to serve as supervision, goal-selection, and evaluation ground truth.
    §IV.A.2 uses AnyTraverse outputs as the teacher signal; §V.F uses the same teacher-derived cost C(x) and map T for metrics. If the teacher is systematically wrong, the simulated evaluation is meaningless.
  • ad hoc to paper Randomized Gaussian waypoint sampling + Bézier smoothing + regularizers yields a training distribution that covers feasible, collision-free trajectories.
    This is the planner-free supervision premise introduced in §IV.A.1 and §IV.B. The paper provides no proof or independent evidence that these synthetic trajectories adequately represent the space of safe paths.
  • domain assumption Projection of an image-space trajectory onto the point cloud, plus small APF corrections, yields a physically executable 3D path.
    The paper acknowledges 'mild geometric uncertainty' in §I and relies on APF refinement for real-world safety; the magnitude of APF corrections is not ablated.
  • domain assumption DINOv2 visual features, modulated by FiLM state vectors, transfer across indoor scenes and robot embodiments.
    The student model freezes DINOv2 (§IV.A.3) and assumes its features are embodiment-modulatable without fine-tuning; no analysis of feature failure modes is provided.
  • standard math DDPM noise prediction with a cosine schedule is a valid training objective for sparse, multimodal trajectory heatmaps.
    This is a standard diffusion assumption; the paper argues empirically for noise prediction over sample prediction in §V.G.4, but the justification is empirical rather than derived.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EmbodiedDiffusion: End-to-End Traversability-Guided Visual Diffusion for Heterogeneous Robot Navigation." pith.science (2026). https://pith.science/paper/NZFD63SL

@misc{pith2026251202851,
  author       = {Pith},
  title        = {Pith review of: EmbodiedDiffusion: End-to-End Traversability-Guided Visual Diffusion for Heterogeneous Robot Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NZFD63SL}},
  note         = {Machine review of arXiv:2512.02851}
}
read the original abstract

Visual traversability estimation is central to autonomous navigation, yet most approaches either rely on prompt-driven Vision-Language Model (VLM) or decouple traversability from trajectory planning, requiring separate planners with heavy mapping, manual tuning, and extended deployment time. We propose EmbodiedDiffusion, a diffusion-based framework that simultaneously predicts traversability maps and generates feasible trajectories from RGB images using planner-free synthetic supervision and embodiment conditioning for cross-platform transfer. The framework distills category-level traversability semantics from a VLM teacher into a lightweight student model during training, enabling prompt-free, real-time inference at deployment. A modular FiLM-based conditioning mechanism isolates embodiment-specific reasoning into a compact trainable subset of the network, allowing rapid adaptation to new robot platforms without retraining the visual backbone or the trajectory diffusion model. Across indoor environments with quadruped and aerial robots, EmbodiedDiffusion achieves 80-100% navigation success in the full-data regime with real-time inference (90 ms) and adapts to new platforms using only 10 min of visual data collection, demonstrating scalable, unified traversability reasoning and trajectory generation for heterogeneous robots.

Figures

Figures reproduced from arXiv: 2512.02851 by the authors.

Figure 1
Figure 1. Cross-embodiment traversability transfer and feasible trajectory [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The proposed model consists of two interconnected components: (1) [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. UAV simulation architecture. For this work, we developed a simulated environment. An overview of the system is shown in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison between baseline and our method across drone and quadruped platforms. Rows show (i) baseline trajectories, (ii) baseline [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 2 linked inside Pith

  1. [1]

    A smart robotic system for industrial plant supervision,

    D. A. G ´omez-Rosal, M. Bergau, G. K. Fischer, A. Wachaja, J. Grater, M. Odenweller, U. Piechottka, F. Hoeflinger, N. Gosala, N. Wetzel, D. Buetscher, A. Valada, and W. Burgard, “A smart robotic system for industrial plant supervision,” in2023 IEEE SENSORS, pp. 1–4, 2023

    2023

  2. [2]

    Ad- ventr: Autonomous robot navigation in complex outdoor environments,

    K. Weerakoon, A. J. Sathyamoorthy, M. Elnoor, and D. Manocha, “Ad- ventr: Autonomous robot navigation in complex outdoor environments,” inInternational Symposium on Experimental Robotics, pp. 219–228, Springer, 2023

    2023

  3. [3]

    Fast Traversability Estimation for Wild Visual Navigation,

    J. Frey, M. Mattamala, N. Chebrolu, C. Cadena, M. Fallon, and M. Hutter, “Fast Traversability Estimation for Wild Visual Navigation,” inProceedings of Robotics: Science and Systems, (Daegu, Republic of Korea), July 2023. 12

    2023

  4. [4]

    Wayfaster: a self-supervised traversability prediction for increased navigation aware- ness,

    M. V . Gasparino, A. N. Sivakumar, and G. Chowdhary, “Wayfaster: a self-supervised traversability prediction for increased navigation aware- ness,” inProc. IEEE Int. Conf. on Rob. and Auto., pp. 8486–8492, IEEE, 2024

    2024

  5. [5]

    Anytraverse: An off-road traversability framework with vlm and human operator in the loop,

    S. Sahu, A. Singh, K. Nambiar, S. Saripalli, and P. Sujit, “Anytraverse: An off-road traversability framework with vlm and human operator in the loop,”arXiv preprint arXiv:2506.16826, 2025

    arXiv 2025

  6. [6]

    Interactive navigation in environments with traversable obstacles using large language and vision-language models,

    Z. Zhang, A. Lin, C. W. Wong, X. Chu, Q. Dou, and K. W. Samuel Au, “Interactive navigation in environments with traversable obstacles using large language and vision-language models,” in2024 IEEE International Conference on Robotics and Automation (ICRA), pp. 7867–7873, 2024

    2024

  7. [7]

    V-strong: Visual self-supervised traversability learning for off-road navigation,

    S. Jung, J. Lee, X. Meng, B. Boots, and A. Lambert, “V-strong: Visual self-supervised traversability learning for off-road navigation,” inProc. IEEE Int. Conf. on Rob. and Auto., pp. 1766–1773, 2024

    2024

  8. [8]

    Badgr: Learning to drive from raw sensor data using self-supervision,

    R. Julian, K. Stone, W. Koch, A. Faust, I. Gilitschenski, and J. P. How, “Badgr: Learning to drive from raw sensor data using self-supervision,” inProc. Rob.: Sci. and Syst., 2020

    2020

Show all 37 references
  1. [9]

    Learning- on-the-drive: Self-supervised adaptation of visual offroad traversability models,

    E. Chen, C. Ho, M. Maulimov, C. Wang, and S. Scherer, “Learning- on-the-drive: Self-supervised adaptation of visual offroad traversability models,” 2024

    2024

  2. [10]

    Adaptive robot traversability estimation based on self-supervised online continual learning in unstructured environments,

    H.-S. Yoon, J.-H. Hwang, C. Kim, E. I. Son, S.-W. Yoo, and S.-W. Seo, “Adaptive robot traversability estimation based on self-supervised online continual learning in unstructured environments,”IEEE Rob. and Auto. Let., vol. 9, no. 6, pp. 4902–4909, 2024

    2024

  3. [11]

    Foresttrav: 3d lidar-only forest traversability estimation for autonomous ground vehicles,

    F. A. Ruetz, N. Lawrance, E. Hern ´andez, P. V . K. Borges, and T. Peynot, “Foresttrav: 3d lidar-only forest traversability estimation for autonomous ground vehicles,”IEEE Access, vol. 12, pp. 37192–37206, 2024

    2024

  4. [12]

    Self-supervised visual terrain classification from unsupervised acoustic feature learning,

    J. Zurn, W. Burgard, and A. Valada, “Self-supervised visual terrain classification from unsupervised acoustic feature learning,”IEEE Trans- actions on Robotics, vol. 37, pp. 466–481, 2019

    2019

  5. [13]

    Salon: Self-supervised adaptive learning for off-road navigation,

    M. Sivaprakasam, S. Triest, C. Ho, S. Aich, J. Lew, I. Adu, W. Wang, and S. Scherer, “Salon: Self-supervised adaptive learning for off-road navigation,”arXiv preprint arXiv:2412.07826, 2024

    2024 arXiv

  6. [14]

    Scene perception based visual navigation of mobile robot in indoor environment,

    T. Ran, L. Yuan, and J. Zhang, “Scene perception based visual navigation of mobile robot in indoor environment,”ISA Transactions, vol. 109, pp. 389–400, 2021

    2021

  7. [15]

    Object-based reliable visual navigation for mobile robot,

    F. Wang, C. Zhang, W. Zhang, C. Fang, Y . Xia, Y . Liu, and H. Dong, “Object-based reliable visual navigation for mobile robot,”Sensors, vol. 22, no. 6, 2022

    2022

  8. [16]

    Vision-based distributed multi-uav collision avoidance via deep reinforcement learning for navigation,

    H. Huang, G. Zhu, Z. Fan, H. Zhai, Y . Cai, Z. Shi, Z. Dong, and Z. Hao, “Vision-based distributed multi-uav collision avoidance via deep reinforcement learning for navigation,” inProc. 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 13745–...

    2022

  9. [17]

    Deep reinforcement learning assisted uav path planning relying on cumulative reward mode and region segmentation,

    Z. Wang, S. X. Ng, and M. EI-Hajjar, “Deep reinforcement learning assisted uav path planning relying on cumulative reward mode and region segmentation,”IEEE Open Journal of Vehicular Technology, vol. 5, pp. 737–751, 2024

    2024

  10. [18]

    Safemove-rl: A certifiable reinforcement learning framework for dynamic motion constraints in trajectory planning,

    T. Liu, H. Zhong, J. Hu, and T. Zhang, “Safemove-rl: A certifiable reinforcement learning framework for dynamic motion constraints in trajectory planning,” 2025

    2025

  11. [19]

    Dream to fly: Model-based reinforcement learning for vision-based drone flight,

    A. Romero, A. Shenai, I. Geles, E. Aljalbout, and D. Scaramuzza, “Dream to fly: Model-based reinforcement learning for vision-based drone flight,” 2025

    2025

  12. [20]

    Dippest: Diffusion-based path planner for synthesizing trajectories applied on quadruped robots,

    M. Stamatopoulou, J. Liu, and D. Kanoulas, “Dippest: Diffusion-based path planner for synthesizing trajectories applied on quadruped robots,” 2024

    2024

  13. [21]

    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 Inter- national Conference on Robotics and Automation (ICRA), pp. 63–70, 2024

    2024

  14. [22]

    Motion planning diffusion: Learning and adapting robot motion planning with diffusion models,

    J. Carvalho, A. T. Le, P. Kicki, D. Koert, and J. Peters, “Motion planning diffusion: Learning and adapting robot motion planning with diffusion models,”IEEE Transactions on Robotics, vol. 41, pp. 4881–4901, 2025

    2025

  15. [23]

    Ventura: Adapting image diffusion models for unified task conditioned navigation,

    A. Zhang, X. Meng, L. Calliari, D.-K. Kim, S. Omidshafiei, J. Biswas, A. Agha, and A. Shaban, “Ventura: Adapting image diffusion models for unified task conditioned navigation,”arXiv preprint arXiv:2510.01388, 2025

    2025

  16. [24]

    Navidiffusor: Cost-guided diffusion model for visual navigation,

    Y . Zeng, H. Ren, S. Wang, J. Huang, and H. Cheng, “Navidiffusor: Cost-guided diffusion model for visual navigation,” in2025 IEEE Inter- national Conference on Robotics and Automation (ICRA), pp. 11994– 12001, 2025

    2025

  17. [25]

    Flightdiffusion: Revolutionising autonomous drone training with diffusion models generating fpv video,

    V . Serpiva, A. Lykov, F. Batool, V . Kozlovskiy, M. A. Cabrera, and D. Tsetserukou, “Flightdiffusion: Revolutionising autonomous drone training with diffusion models generating fpv video,” 2025

    2025

  18. [26]

    Gps denied ibvs-based navigation and collision avoidance of uav using a low-cost rgb camera,

    X. Wang, Y . R. Tan, W. Leong, S. Huang, R. Teo, and C. Xiang, “Gps denied ibvs-based navigation and collision avoidance of uav using a low-cost rgb camera,” 2025

    2025

  19. [27]

    EAST: environment-aware safe tracking for robot navigation in dynamic environments,

    Z. Li, Y . Yi, Z. Niu, and N. Atanasov, “EAST: environment-aware safe tracking for robot navigation in dynamic environments,”Autonomous Robots, vol. 49, no. 4, p. 36, 2025

    2025

  20. [28]

    Neuroswarm: Multi-agent neural 3d scene reconstruc- tion and segmentation with uav for optimal navigation of quadruped robot,

    I. Zhura, D. Davletshin, N. D. W. Mudalige, A. Fedoseev, R. Peter, and D. Tsetserukou, “Neuroswarm: Multi-agent neural 3d scene reconstruc- tion and segmentation with uav for optimal navigation of quadruped robot,” in2023 IEEE International Conference on Systems, Man, and Cybe...

    2023

  21. [29]

    Hma-sar: Multi-agent search and rescue for unknown located dynamic targets in completely unknown environments,

    X. Cao, M. Li, Y . Tao, and P. Lu, “Hma-sar: Multi-agent search and rescue for unknown located dynamic targets in completely unknown environments,”IEEE Robotics and Automation Letters, vol. 9, no. 6, pp. 5567–5574, 2024

    2024

  22. [30]

    Ventura: Adapting image diffusion models for unified task conditioned navigation,

    A. Zhang, X. Meng, L. Calliari, D.-K. Kim, S. Omidshafiei, J. Biswas, A. Agha, and A. Shaban, “Ventura: Adapting image diffusion models for unified task conditioned navigation,” 2025

    2025

  23. [31]

    DINOv2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V . V o, M. Szafraniec, V . Khali- dov, P. Fernandez, D. HAZIZA, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P.-Y . Huang, S.-W. Li, I. Misra, M. Rabbat, V . Sharma, G. Synnaeve, H. Xu, H. Jegou, J. Mairal, P. Labatu...

    2024

  24. [32]

    Vision transformer adapter for dense predictions,

    Z. Chen, Y . Duan, W. Wang, J. He, T. Lu, J. Dai, and Y . Qiao, “Vision transformer adapter for dense predictions,” inInt. Conf. Learn. Represent., 2023

    2023

  25. [33]

    Film: Visual reasoning with a general conditioning layer,

    E. Perez, F. Strub, H. De Vries, V . Dumoulin, and A. Courville, “Film: Visual reasoning with a general conditioning layer,” inProceedings of the AAAI conference on artificial intelligence, vol. 32, 2018

    2018

  26. [34]

    Diffusers: State-of-the-art diffusion models

    P. von Platen, S. Patil, A. Lozhkov, P. Cuenca, N. Lambert, K. Rasul, M. Davaadorj, D. Nair, S. Paul, W. Berman, Y . Xu, S. Liu, and T. Wolf, “Diffusers: State-of-the-art diffusion models.” https://github. com/huggingface/diffusers, 2022

    2022

  27. [35]

    Flightmare: A flexible quadrotor simulator,

    Y . Song, S. Naji, E. Kaufmann, A. Loquercio, and D. Scaramuzza, “Flightmare: A flexible quadrotor simulator,” inConference on Robot Learning, 2020

    2020

  28. [36]

    Openvins: A research platform for visual-inertial estimation,

    P. Geneva, K. Eckenhoff, W. Lee, Y . Yang, and G. Huang, “Openvins: A research platform for visual-inertial estimation,” in2020 IEEE Interna- tional Conference on Robotics and Automation (ICRA), pp. 4666–4672, 2020

    2020

  29. [37]

    DTG : Diffusion-based Trajectory Generation for Mapless Global Navigation,

    J. Liang, A. Payandeh, D. Song, X. Xiao, and D. Manocha, “DTG : Diffusion-based Trajectory Generation for Mapless Global Navigation,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 5340–5347, 2024

    2024

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.