Pith. sign in

REVIEW 4 major objections 5 minor 22 references

Learning to Navigate Efficiently with Only 0.58M Trainable Parameters

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a decomposed planner with only 0.58M trainable parameters can approach state-of-the-art point-goal navigation by computing geometry analytically instead of learning it.

desk verdict A genuinely interesting efficiency result built on a clean decomposition and honest ablations, but the missing statistical rigor and an underspecified egress supervision edge case keep it from being a slam dunk. read the letter →

arxiv 2607.11029 v2 pith:2CAWW4JG submitted 2026-07-13 cs.RO cs.CV

Edward Beng Wai Tan , Siew-Kei Lam This is my paper · ORCID
classification cs.ROcs.CV
keywords point-goalnavigationdecomposedpolicyegresspredictiondiffusiontrajectorygenerationB-splineparameterizationBEVoccupancyefficientlearningcollisionavoidance
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 asks how much of the massive scale behind recent visual-navigation policies is actually needed for a single task family, point-goal navigation. Its answer is a decomposition: three small learned modules — an egress predictor that grounds the goal in the current view, a navigation predictor that estimates where trajectories travel, and an endpoint-pinned residual diffusion generator that shapes trajectories — connected by closed-form projective geometry and occupancy transforms. Because the operations with known structure are computed analytically, the learned part shrinks to 0.58M trainable parameters (out of 22.7M, the rest a frozen image encoder), trained on 44k frames in under one GPU-hour. The paper reports near state-of-the-art success on 6060 point-goal episodes across 60 environments, the lowest collision rate among evaluated methods, and 10+ Hz inference on a Jetson Xavier NX. A sympathetic reader would care because, if right, it shows that scale is not the only route to capable navigation, and that structure can substitute for data and parameters.

What carries the argument

The central mechanism is the interface between modules: learned operations live in their native coordinate frames (image frame for egress, BEV frame for navigation posterior and generator), and the transforms between them — unprojecting depth to occupancy, ray-casting the egress pixel to a 3D endpoint, B-spline residual parameterization around the straight line to that endpoint — are computed in closed form. The egress predictor's image-plane exit point c_t is the load-bearing object: it is where the path to the goal exits the current BEV window or passes behind objects, subject to a distance truncation, and it converts the episode goal into a local, pixel-anchored subgoal. A second key desi

What would settle it

On the 6060-episode benchmark, count how many episodes require a goal behind the robot or outside the camera's horizontal field of view at any step; if such states occur and the egress predictor still succeeds, its behavior on those states is determined by an unspecified boundary-clamping mechanism, and modifying or ablating that mechanism should measurably change success rate. Concretely: run the released model with the egress argmax masked to exclude pixels outside the view frustum and measure the success-rate drop.

Watch

Extended reading notes

Core claim

The central claim is that point-goal visual navigation is realizable as a composition of three sub-tasks, each learned by a small operator, with analytic geometry serving as the interface: the egress predictor maps the episode goal to a pixel subgoal in the current image; the navigation predictor maps the depth-derived BEV occupancy and that subgoal to a goal-conditioned posterior over where trajectories travel; and the generator, a B-spline residual diffusion model, samples trajectory shapes from this posterior, pinned at the egress endpoint. Training only the 0.58M parameters of these modules — the image encoder stays frozen — on 44k frames for under one GPU-hour yields near-SOTA performan

Load-bearing premise

The egress predictor's training label is an image-plane point where the path to the goal exits the current BEV window or passes behind objects, which presumes the goal path always projects to a pixel in the forward camera view; the paper never states how goals behind the robot or outside the field of view are labeled, truncated, or excluded.

Editorial extensions

If this is right

  • If the decomposition is right, near-SOTA point-goal navigation is attainable with 233x fewer trainable parameters than the 135.7M-parameter diffusion policy, making on-device inference at 10 Hz practical on edge GPUs.
  • The egress interface is reusable: retraining only the 123k-parameter egress head adapts the same system to no-goal exploration, outperforming the large policy's native exploration time in the paper's 100-trial evaluation.
  • Because geometry is computed analytically, sensor failures are localized and correctable: heavy Gaussian depth noise breaks the BEV projection, and a 3x3 median filter restores baseline success — a transparency property end-to-end policies lack.
  • Training cost drops to under one GPU-hour on 44k frames, suggesting that data and compute budgets for a single task family can be orders of magnitude smaller than the billion-frame regime.
  • The ablation results imply the navigation posterior is indispensable: removing it collapses success to 3.0% SR, and joint end-to-end training similarly degrades performance to 35.6% SR.

Reading between the lines

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

  • The paper's single-floor limitation suggests a direct extension: replacing the BEV occupancy with a voxel occupancy would let the same three-module decomposition handle multi-floor navigation, with the egress predictor selecting a subgoal in 3D instead of a pixel.
  • The egress predictor's image-plane supervision has an implicit boundary condition for goals outside the camera FOV; a testable extension is to explicitly define a 'no-egress' class or a clamping rule and measure how often such states occur in the 6060-episode benchmark.
  • Because the generator never sees raw sensor input, one could attach the same navigation posterior and generator to different perception front-ends (e.g., learned depth from monocular RGB), isolating perception quality from planning quality.
  • The posterior-as-target training scheme suggests a curriculum: annealing from a straight-line prior to the ground-truth route could be read as a form of progressive refinement; one could test whether more than the K refinement steps improve the posterior's corridor fidelity.
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

4 major / 5 minor

Summary. The paper proposes a decomposed point-goal visual navigation system: an egress predictor maps the episode goal to an image-plane subgoal, a navigation predictor estimates a BEV goal-conditioned posterior over trajectory locations, and an endpoint-pinned residual diffusion generator samples trajectory shapes. Projective geometry, occupancy, and coordinate transforms are computed analytically as interfaces. The system uses only 0.58M trainable parameters (out of 22.7M including a frozen ViT encoder), trains on 44k frames in under one GPU-hour, and reports near state-of-the-art SR/SPL on the InternRobotics point-goal benchmarks (6060 episodes, 60 environments), the lowest collision rate among compared methods, and 10+ Hz inference on a Jetson Xavier NX. Ablations identify each module as load-bearing and show transfer to no-goal exploration by retraining only the 123k-parameter egress head.

Significance. If the reported results are reliable, the paper makes a strong case that explicit geometric decomposition can substitute for large-scale end-to-end learning in point-goal navigation, with 233x fewer trainable parameters and substantially lower collision rates. The paper has concrete strengths: comparisons against three recent baselines using publicly released weights, a clear ablation structure showing each module matters, a sensor-corruption study with analytically explainable failure modes, and an efficiency evaluation on an embedded GPU. The central claim is falsifiable and measured on a held-out benchmark, not derived from the model's own fitted values. However, the experimental reporting currently lacks uncertainty quantification, and one load-bearing training-signal definition is underspecified, so the significance is contingent on the authors resolving these issues.

major comments (4)
  1. [§III-B, Eq. (3)] The egress training target c_t is defined as 'the point on the rasterized image-frame where the path to Ĝ_t exits the current BEV window or passes behind objects, subject to a maximum distance truncation.' This presupposes that the path's exit projects to a pixel in the forward camera image. For goals behind the robot or outside the horizontal field of view, no such pixel exists. The paper does not state how these states are labeled (e.g., clamped to the boundary, ignored, or assigned a default), nor how inference-time arg max is interpreted for such states. Since the egress module is the sole mechanism that grounds the global goal as a local subgoal — and Table V shows removing it drops SR from 58.4 to 20.8 — the main benchmark results may depend on an unstated boundary behavior. Please specify the labeling policy, report the fraction of benchmark states in which c_t is undefined, and s
  2. [§IV-B, Tables I–II] No confidence intervals, standard errors, or number of evaluation seeds are reported for the main results. The closest comparison, Home SR (Ours 51.4 vs NavDP 57.2, SPL 45.8 vs 52.2), is a gap that could plausibly be within sampling noise if per-environment variation is large. The paper's central claim of 'approaching SOTA' and the collision-rate advantage deserve statistical support. Please report variance across episodes/environments, or run multiple seeds and report mean±std. This is particularly important for the 2–6% SR differences in Table I and for the collision-rate comparison in Table II.
  3. [§IV-E, Tables V–VI] The ablations are run on '101 episodes' but the subset is not characterized. The full system attains SR 58.4 on this subset (Table V), whereas the full InternScenes-Home split in Table I shows SR 51.4 for the same configuration. This nearly 7-point discrepancy suggests the 101-episode subset is not representative of the Table I Home split, or the evaluation protocol differs. Since the ablations are used to claim that each component is load-bearing, please clarify the subset selection, report its difficulty relative to the full benchmark, and, if the subset is easier, temper the quantitative claims or re-run on the full set.
  4. [§III-D and Table VI] The generator is trained on 'the pre-computed f_np posterior predictions rather than the ground-truth' (Sec. III-D), and Table VI shows that training on ground-truth trajectories degrades SR (34.7 vs 58.4). This is a self-referential training loop in which the generator learns from the navigation predictor's own posterior. While the ablation indicates this is beneficial, the mechanism is not explained: why would training against a flawed posterior generalize better than training against ground-truth trajectories, given that the posterior is itself learned? Please provide an analysis (e.g., multimodality preservation) and, ideally, an experiment that varies the quality of the posterior to show the effect is not an artifact of the 101-episode subset.
minor comments (5)
  1. [Fig. 2] The caption mentions a '2D Goal Critic', but the architecture diagram and text do not describe a learned critic; Section III-E describes a post-hoc deterministic selection over candidates. Please align the figure terminology with the method.
  2. [§IV-E] Typo: 'softarg max' should be 'soft argmax' or 'soft-argmax'.
  3. [§IV-D] The claim of 'above 10 Hz on Jetson Xavier NX' is stated without measurement details (e.g., image resolution, batch size, input/output tensors, or whether the image encoder is included). Please provide a short benchmark description for reproducibility.
  4. [Eqs. (7)–(8)] The same symbol λ_c is used for the collision penalty in the navigation-predictor loss and in the generator loss, but the two terms are computed on different objects (posterior map vs. trajectory collocation). Please clarify whether they share a value or are separate hyperparameters.
  5. [Table III] Reporting 'relative SR (%)' with clean baseline set to 100 is helpful, but the absolute clean SR is not given for the 101-episode Home subset. Please include absolute numbers so readers can judge the effect size.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; all learned modules are supervised by external ground-truth path geometry and evaluated on held-out benchmarks.

full rationale

No equation-level circularity found. Each learnable module has an external supervision target: the egress predictor is trained against the ray-traced exit point c_t (Eq. 3), the navigation posterior against an annealed ground-truth route tau_k (Eq. 5), and the generator against posterior samples plus a geometric collision cost (Eq. 8). The final trajectory selection (Eq. 9) optimizes a costmap proxy, but the reported SR, SPL, and collision metrics are measured on held-out InternRobotics episodes against external baselines using their released weights. The paper contains no load-bearing self-citations; its citations to SanD-Planner, DDPM, and NavDP are external architectural or baseline references. The reviewer-identified gap that the egress label c_t is undefined for goals outside the camera FOV (Sec. III-B) is a specification/correctness concern, not a circular reduction: it does not make any prediction equivalent to its own input by construction. The paper's own stated limitations (single-floor BEV maps, simulation-only evaluation) similarly affect generality but not circularity. The central result is an empirical comparison, not a derivation forced by the model's own fitted values.

Assumptions & free parameters 10 free parameters · 8 assumptions · 0 invented entities

The central result depends on hand-set hyperparameters (costmap scales, loss weights, annealing schedule, grid resolutions), standard geometry assumptions, and three unstated domain assumptions (training/eval disjointness, goal visibility in the camera frame, single-frame Markov state). The learned 0.58M module weights are trained, not hand-fit, and are therefore not enumerated as free parameters; no physical invented entities are introduced.

free parameters (10)
  • λ_d (costmap distance-decay scale)
    Appears in Eq 1 and Eq 9; controls how strongly obstacles repel the costmap and the critic; value not stated.
  • ρ(x) (bearing ramp in costmap)
    Added in Eq 1 to penalize out-of-view mass; functional form and parameters not defined.
  • λ_c (collision penalty weight)
    Used in L_np (Eq 7) and L_gen (Eq 8) to weight collision cost; value not stated; directly shapes the low-collision-rate result.
  • λ_θ (heading loss weight)
    Weights tangent-direction term in Eq 6; value not stated.
  • α_k annealing schedule = k/K (Eq 5)
    Hand-chosen linear schedule interpolating the nav-posterior supervision target from straight-line to ground-truth path; this defines what the 'posterior' must learn.
  • B-spline interior control points M
    Residual trajectory parameterization δ ∈ R^{M×2}; M not stated.
  • Diffusion denoising steps = 10
    Generator trains/denoises over 10 steps (Sec III-D); noise schedule and β values not given.
  • Candidate trajectory count N
    Number of sampled paths in the generator and critic selection (Sec III-D/E); not stated.
  • Egress max-distance truncation
    The exit-point label c_t is 'subject to a maximum distance truncation' (Sec III-B); truncation radius not specified.
  • BEV window size/resolution
    Occupancy, costmap, and nav posterior operate on a rasterized BEV grid; resolution and extent not stated.
assumptions (8)
  • standard math Pinhole camera projection, depth unprojection, and ray-casting are exact enough to compute BEV occupancy and 3D subgoal endpoints.
    Used in Sec III-A/B to derive costmap O_t and endpoint g_t; standard geometry, not verified in paper.
  • standard math B-spline parameterization provides C² continuity and the residual form τ=τ_0(gt)+Bδ.
    Borrowed from SanD-Planner [11], Sec III-D; not re-derived.
  • domain assumption A depth-derived BEV occupancy map is a sufficient state representation for single-floor point-goal navigation.
    Sec III-A/V; fails on multi-floor or elevation changes; authors acknowledge restriction.
  • domain assumption The egress exit-point label c_t exists as a pixel for every training and evaluation state.
    Sec III-B defines c_t as image-frame exit point; goals behind/outside FOV have no such pixel; no fallback stated.
  • domain assumption The current frame (single RGB+Depth) is sufficient to predict the subgoal and trajectory; no memory of past frames is needed.
    All modules operate per frame (Sec III-A); if global context is required, the Markov assumption breaks.
  • ad hoc to paper The annealing schedule τ_k = (1-α_k)τ_prior + α_k τ_gt (Eq 5) produces meaningful supervision for the navigation posterior.
    Introduced by this paper as a training target; not grounded in a physical or perceptual principle; central to nav-predictor training.
  • ad hoc to paper Training the diffusion generator on samples from the learned navigation posterior, instead of ground-truth trajectories, yields a better policy.
    Sec III-D and Table VI ablation support it empirically within this paper; it is a self-distillation assumption not independently established.
  • domain assumption InternRobotics evaluation environments are disjoint from the 55 Matterport3D training scenes.
    Sec IV-A/IV-B: training scenes and eval environments are described separately but overlap is never ruled out; the generalization claim depends on disjointness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning to Navigate Efficiently with Only 0.58M Trainable Parameters." pith.science (2026). https://pith.science/paper/2CAWW4JG

@misc{pith2026260711029,
  author       = {Pith},
  title        = {Pith review of: Learning to Navigate Efficiently with Only 0.58M Trainable Parameters},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2CAWW4JG}},
  note         = {Machine review of arXiv:2607.11029}
}
abstract

Recent progress in visual navigation has largely been driven by scale: end-to-end policies with hundreds of millions of parameters trained on billions of frames or large-scale simulated data. We ask how much of this scale a single task family actually requires, and what structure can substitute for it. We propose a decomposed navigation model in which operations with known closed-form structure, such as projective geometry, occupancy, and coordinate transforms, are computed analytically and serve as interfaces between three small learned modules: an egress predictor that grounds the episode goal as a local subgoal in the current view, a navigation predictor that estimates a goal-conditioned posterior over where trajectories travel, and an endpoint-pinned residual diffusion generator that samples trajectory shapes from this posterior. The system trains only 0.58M out of a total of 22.7M parameters, on 44k frames in under one GPU-hour, yet approaches the performance of state-of-the-art models on navigation tasks across 6060 point-goal episodes and 60 environments, while having 233$\times$ fewer trainable parameters, the lowest collision rate among all evaluated methods, and 10+ Hz inference rate on a Jetson Xavier NX. The decomposition further transfers to no-goal exploration by retraining only the 123k-parameter egress head, and its failure modes under sensor corruption are transparent and analytically correctable.

Figures

Figures reproduced from arXiv: 2607.11029 by the authors.

Figure 1
Figure 1. Comparison of navigation paradigms. (a) Classical navigation algorithms are efficient but hand-designed heuris￾tics can limit performance, (b) end-to-end navigation models learn geometry, mapping, and control implicitly, at the cost of large parameter and scaling cost, (c) our method retains closed￾form solutions for operations with known structure, allowing the network to learn the rest of the navigation task with … view at source ↗
Figure 2
Figure 2. System Architecture. Our system consists of three learnable modules, each performing a distinct task. The egress predictor feg learns a local navigation subgoal, the navigation predictor fnp learns goal-conditioned spatial feasibility, and the generator fgen predicts the trajectory shape between the current position and predicted subgoal. Modules operate in their native coordinate frames and exchange information thr… view at source ↗
Figure 4
Figure 4. Qualitative comparison showing ours successfully navigating difficult conditions. Areas indicated in red denote episode failure points. In (a), suboptimal initialization position causes ViPlanner and iPlanner to collide with the wall, in (b) all others fail to navigate the tight space, and in (c) NavDP takes a suboptimal route, resulting in too sharp of an approach. marginally worse: and our findings are that in mos… view at source ↗

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 6 linked inside Pith

  1. [1]

    DD-PPO: Learning near-perfect PointGoal navigators from 2.5 billion frames,

    E. Wijmans, A. Kadian, A. Morcos, S. Lee, I. Essa, D. Parikh, M. Savva, and D. Batra, “DD-PPO: Learning near-perfect PointGoal navigators from 2.5 billion frames,” inInternational conference on learning representations, 2020. [Online]. Available: https://openreview.net/forum?id=H1gX8C4YPr

    2020

  2. [2]

    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. [Online]. Available: https://arxiv.org/abs/2210.03370

    arXiv 2023

  3. [3]

    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. [Online]. Available: https://arxiv.org/abs/2306.14846

    arXiv 2023

  4. [4]

    NavDP: Learning Sim-to-Real Navigation Diffusion Policy with Privileged Information Guidance,

    W. Cai, J. Peng, Y . Yang, Y . Zhang, M. Wei, H. Wang, Y . Chen, T. Wang, and J. Pang, “NavDP: Learning Sim-to-Real Navigation Diffusion Policy with Privileged Information Guidance,” 2025, version Number: 3. [Online]. Available: https://arxiv.org/abs/2505.08712

    arXiv 2025

  5. [5]

    LoGoPlanner: Localization Grounded Navigation Policy with Metric- aware Visual Geometry,

    J. Peng, W. Cai, Y . Yang, T. Wang, Y . Shen, and J. Pang, “LoGoPlanner: Localization Grounded Navigation Policy with Metric- aware Visual Geometry,” 2025, version Number: 2. [Online]. Available: https://arxiv.org/abs/2512.19629

    arXiv 2025

  6. [6]

    Diffusion policy: Visuomotor policy learning via action diffusion,

    C. Chi, S. Feng, Y . Du, Z. Xu, E. Cousineau, B. Burchfiel, and S. Song, “Diffusion policy: Visuomotor policy learning via action diffusion,” in Proceedings of robotics: Science and systems (RSS), 2023

    2023

  7. [7]

    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,”arXiv pre-print,

  8. [8]

    Prior does matter: Visual navigation via denoising diffusion bridge models,

    H. Ren, Y . Zeng, Z. Bi, Z. Wan, J. Huang, and H. Cheng, “Prior does matter: Visual navigation via denoising diffusion bridge models,” in Proceedings of the computer vision and pattern recognition conference, 2025, pp. 12 100–12 110

    2025

Show all 22 references
  1. [9]

    StepNav: Structured trajectory priors for efficient and multimodal visual navigation,

    X. Luo, A. Wu, H. Han, X. Wan, W. Zhang, L. Shu, and R. Wang, “StepNav: Structured trajectory priors for efficient and multimodal visual navigation,” 2026, arXiv: 2602.02590 [cs.RO]. [Online]. Available: https://arxiv.org/abs/2602.02590

    2026

  2. [10]

    Rectified Schr ¨odinger Bridge Matching for Few-Step Visual Navigation,

    W. Luan, J. Li, W. Zhao, W. Zhang, T. Wu, and R. Ma, “Rectified Schr ¨odinger Bridge Matching for Few-Step Visual Navigation,” May 2026, arXiv:2604.05673 [cs.RO]. [Online]. Available: http://arxiv.org/abs/2604.05673

    2026 arXiv

  3. [11]

    SanD-Planner: Sample-Efficient Diffusion Planner in B-Spline Space for Robust Local Navigation,

    J. Wang, L. Bao, T. Yang, D. M. Plasencia, J. Jiao, and D. Kanoulas, “SanD-Planner: Sample-Efficient Diffusion Planner in B-Spline Space for Robust Local Navigation,” 2026, version Number: 1

    2026

  4. [12]

    The dynamic window approach to collision avoidance,

    D. Fox, W. Burgard, and S. Thrun, “The dynamic window approach to collision avoidance,”IEEE Robotics & Automation Magazine, vol. 4, no. 1, pp. 23–33, Mar. 1997. [Online]. Available: http://ieeexplore.ieee.org/document/580977/

    1997

  5. [13]

    Adaptive and explainable deployment of navigation skills via hierarchical deep reinforcement learning,

    K. Lee, S. Kim, and J. Choi, “Adaptive and explainable deployment of navigation skills via hierarchical deep reinforcement learning,” in 2023 IEEE international conference on robotics and automation (ICRA), 2023, pp. 1673–1679

    2023

  6. [14]

    Object goal navigation using goal-oriented semantic exploration,

    D. S. Chaplot, D. P. Gandhi, A. Gupta, and R. R. Salakhutdinov, “Object goal navigation using goal-oriented semantic exploration,” inAdvances in neural information processing systems, H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, Eds., vol. 33. Curran Associate...

    2020

  7. [15]

    Viplanner: Visual semantic imperative learning for local navigation,

    P. Roth, J. Nubert, F. Yang, M. Mittal, and M. Hutter, “Viplanner: Visual semantic imperative learning for local navigation,” in2024 IEEE international conference on robotics and automation (ICRA). IEEE, 2024, pp. 5243–5249

    2024

  8. [16]

    iPlanner: Imperative path planning,

    F. Yang, C. Wang, C. Cadena, and M. Hutter, “iPlanner: Imperative path planning,” inProceedings of robotics: Science and systems, Daegu, Republic of Korea, Jul. 2023

    2023

  9. [17]

    Ground Slow, Move Fast: A Dual-System Foundation Model for Generalizable Vision-and- Language Navigation,

    M. Wei, C. Wan, J. Peng, X. Yu, Y . Yang, D. Feng, W. Cai, C. Zhu, T. Wang, J. Pang, and X. Liu, “Ground Slow, Move Fast: A Dual-System Foundation Model for Generalizable Vision-and- Language Navigation,” 2025, version Number: 1. [Online]. Available: https://arxiv.org/abs/2512.08186

    2025

  10. [18]

    Depth anything V2,

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao, “Depth anything V2,” inAdvances in neural information processing systems, A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, Eds., vol. 37. Curran Associates, Inc., 2024, pp. 21 8...

    2024

  11. [19]

    Denoising Diffusion Probabilistic Models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising Diffusion Probabilistic Models,” Dec. 2020, arXiv:2006.11239 [cs.LG]. [Online]. Available: http://arxiv.org/abs/2006.11239

    2020 arXiv

  12. [20]

    Matterport3D: Learning from RGB- D data in indoor environments,

    A. Chang, A. Dai, T. Funkhouser, M. Halber, M. Niessner, M. Savva, S. Song, A. Zeng, and Y . Zhang, “Matterport3D: Learning from RGB- D data in indoor environments,”International Conference on 3D Vision (3DV), 2017

    2017

  13. [21]

    On Evaluation of Embodied Navigation Agents,

    P. Anderson, A. Chang, D. S. Chaplot, A. Dosovitskiy, S. Gupta, V . Koltun, J. Kosecka, J. Malik, R. Mottaghi, M. Savva, and A. R. Zamir, “On Evaluation of Embodied Navigation Agents,” Jul. 2018, arXiv:1807.06757 [cs.AI]. [Online]. Available: http://arxiv.org/abs/1807.06757

    2018 arXiv

  14. [2023]

    Available: https://arxiv.org/abs/2310.07896

    [Online]. Available: https://arxiv.org/abs/2310.07896

    arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.