pith. sign in

arxiv: 2606.14776 · v2 · pith:NQOV5MFQnew · submitted 2026-06-10 · 💻 cs.RO · cs.LG

Deep Learning-Based Lunar Crater Terrain Relative Navigation

Batu Candan , Simone Servadio This is my paper

Pith reviewed 2026-06-27 10:03 UTC · model grok-4.3

classification 💻 cs.RO cs.LG
keywords lunar crater detectionterrain relative navigationdeep learningextended Kalman filtermonocular imageslunar landingcrater matchingposition estimation
0
0 comments X

The pith

A deep learning crater detector matched to a database and fed into an EKF recovers lunar spacecraft position to a few hundred meters even from a 5 km initial error.

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

The paper develops a terrain relative navigation system for future lunar landings that processes monocular orbital images with a deep learning crater detector. Detected craters are matched to entries in a global database through assignment and outlier removal steps, then supplied as measurements to an extended Kalman filter that also receives altitude data to limit radial error growth. The method targets environments where terrain features are sparse and initial position knowledge may be poor. Simulation results show the filter can bring navigation error down to hundreds of meters when the spacecraft begins up to 5 km off its true location. The approach requires that image resolution and scene scale stay consistent with the detector's training data to keep the crater correspondences reliable.

Core claim

Our deep-learning crater detector analyzes crater features from monocular images acquired from orbit, their matches with craters from a global database are identified via a Hungarian assignment approach followed by the consensus-based outliers removal method, and the estimated measurements are then used to refine an EKF where spacecraft pose estimation in the Lunar-Centered Lunar-Fixed frame of reference, augmented with altitude aiding information, constrains radial drift, achieving navigation error reduction to a few hundred meters from initial errors up to 5 km in simulations.

What carries the argument

Deep-learning crater detector whose outputs are matched to a global database by Hungarian assignment plus consensus outlier removal and then supplied to an extended Kalman filter augmented with altitude aiding in the lunar-centered lunar-fixed frame.

If this is right

  • Spacecraft can obtain usable position updates during lunar orbit using only monocular cameras and a pre-existing crater catalog.
  • Initial position uncertainty of several kilometers can be reduced without external navigation aids.
  • Altitude measurements combined with crater matches constrain the radial component of the navigation error.
  • The system operates in the lunar-centered lunar-fixed frame and therefore produces estimates directly usable for landing guidance.

Where Pith is reading between the lines

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

  • The same detection-plus-matching pipeline could be retrained for other cratered surfaces such as Mars without changing the EKF structure.
  • Performance would degrade on real flight imagery if the camera parameters or lighting conditions fall outside the training distribution.
  • Adding a second sensor modality such as lidar range would relax the strict requirement that image scale exactly match the training set.

Load-bearing premise

The images supplied to the detector must have the same resolution and scale distribution as the images used during training, otherwise crater correspondences cannot be maintained reliably.

What would settle it

Apply the full pipeline to a set of orbital images whose resolution or crater sizes differ substantially from the training distribution and check whether the reported position error stays above several kilometers instead of dropping to hundreds of meters.

Figures

Figures reproduced from arXiv: 2606.14776 by Batu Candan, Simone Servadio.

Figure 2
Figure 2. Figure 2: Representative qualitative result from the NASA Crater Detection [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Representative qualitative result from the NASA Crater Detection [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Data association and quadrant spread logic in the image plane. Red [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Example of crater detection and feature association on a simulated [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: Example synthetic lunar image patch and its corresponding crater [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 8
Figure 8. Figure 8: Monte Carlo 3D position error (N = 500 runs) [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: illustrates the corresponding velocity estimation per￾formance, where the mean 3D velocity error rapidly stabilizes near 2.5 m/s. It is critical to note that this convergence is achieved in a strictly autonomous, bearing-only configuration without the aid of a radar altimeter. By coupling the superior sub-pixel resolving power of the U-Net++ detector with a robust quadrant-spread data association strategy,… view at source ↗
read the original abstract

Accurate position estimation is crucial for the successful implementation of future lunar landings using autonomous vehicles, especially in dangerous environments with sparse terrain features. In this paper, we propose a terrain relative navigation (TRN) algorithm combining our deep-learning crater detector, which was designed specifically for the NASA Crater Detection Challenge problem, and an Extended Kalman Filter (EKF). Our detector analyzes crater features from the monocular images acquired from orbit, and their matches with craters from a global database are identified via a Hungarian assignment approach followed by the consensus-based outliers removal method. The estimated measurements are then used to refine an EKF, where spacecraft pose estimation in the Lunar-Centered Lunar-Fixed (LCLF) frame of reference, augmented with altitude aiding information, constrains radial drift. The simulation results indicate that even if the spacecraft is off from its actual location up to 5 km, TRN could recover from this situation, achieving navigation error reduction to a few hundred meters. It should be noted that in order to maintain crater feature correspondences, it is important to match the image resolution and the scales within the scene to the detector training set distribution.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript presents a terrain-relative navigation (TRN) algorithm for lunar spacecraft that combines a deep-learning crater detector (trained for the NASA Crater Detection Challenge) with an Extended Kalman Filter (EKF). Crater detections from monocular orbital images are matched to a global database using the Hungarian algorithm followed by consensus-based outlier removal; the resulting measurements update an EKF whose state is expressed in the Lunar-Centered Lunar-Fixed frame and is augmented with altitude information to constrain radial drift. The central empirical claim is that the filter can recover from initial position errors of up to 5 km, converging to navigation errors of only a few hundred meters, provided that image scale and resolution remain consistent with the detector’s training distribution.

Significance. Should the reported recovery performance be substantiated by detailed simulation protocols and sensitivity tests, the work would offer a concrete demonstration of learned-feature TRN that could support autonomous lunar landing in regions with sparse natural landmarks. The combination of a challenge-specific detector with standard EKF and assignment methods is a pragmatic engineering contribution.

major comments (2)
  1. [Abstract (simulation-results paragraph)] Abstract (simulation-results paragraph): The headline claim that TRN recovers from a 5 km initial offset to a few hundred meters rests on the EKF receiving usable crater measurements. The text explicitly notes that correspondences require matching image resolution and scene scale to the training distribution, yet the reported simulations supply no information on how scale was controlled or varied when the initial state error is large enough to change apparent crater diameters. Without such detail or an accompanying sensitivity study, it is impossible to assess whether the reported convergence is robust or an artifact of scale-matched rendering.
  2. [Simulation results (throughout)] Simulation results (throughout): No quantitative description is given of the simulation environment, the precise error metrics employed, the number of Monte-Carlo trials, the initial-state covariance, or the validation procedure used to confirm detector performance under the tested conditions. These omissions make the numerical claims difficult to reproduce or compare with prior TRN literature.
minor comments (1)
  1. [Abstract] The phrase 'Lunar-Centered Lunar-Fixed (LCLF) frame of reference' is introduced without a brief definition or reference; a one-sentence clarification would aid readers unfamiliar with lunar coordinate conventions.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major comment below and will revise the manuscript to improve the description of the simulation setup and results.

read point-by-point responses

  1. Referee: The headline claim that TRN recovers from a 5 km initial offset to a few hundred meters rests on the EKF receiving usable crater measurements. The text explicitly notes that correspondences require matching image resolution and scene scale to the training distribution, yet the reported simulations supply no information on how scale was controlled or varied when the initial state error is large enough to change apparent crater diameters. Without such detail or an accompanying sensitivity study, it is impossible to assess whether the reported convergence is robust or an artifact of scale-matched rendering.

    Authors: We agree that the current manuscript lacks explicit details on scale control during simulations with large initial errors. In the revised version, we will expand the simulation description to explain the rendering process used to maintain consistency with the detector training distribution and will add a sensitivity study on scale variations. revision: yes

  2. Referee: No quantitative description is given of the simulation environment, the precise error metrics employed, the number of Monte-Carlo trials, the initial-state covariance, or the validation procedure used to confirm detector performance under the tested conditions. These omissions make the numerical claims difficult to reproduce or compare with prior TRN literature.

    Authors: We acknowledge these omissions limit reproducibility. The revised manuscript will include a detailed description of the simulation environment, definitions of the error metrics, the number of Monte-Carlo trials, the initial-state covariance, and the detector validation procedure under the tested conditions. revision: yes

Circularity Check

0 steps flagged

No circularity; standard detector + EKF pipeline with explicit scale-matching assumption

full rationale

The paper applies a pre-trained deep-learning crater detector to monocular images, uses Hungarian assignment plus consensus for correspondences, and feeds measurements into an EKF augmented with altitude aiding. The reported simulation recovery from 5 km offset is presented as an empirical outcome conditional on the explicitly stated requirement that image resolution and scene scales match the detector's training distribution. No equations or steps reduce a claimed prediction to a fitted parameter by construction, no uniqueness theorems are imported via self-citation, and no ansatz is smuggled in. The central result is a simulation demonstration under a disclosed assumption rather than a derivation that is definitionally equivalent to its inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim relies on the performance of the DL detector and the validity of the EKF model in simulation. No free parameters explicitly mentioned. Reviewed from abstract only.

axioms (2)
  • domain assumption Extended Kalman Filter assumptions hold for the lunar navigation dynamics
    The paper uses EKF without discussing linearization errors or model mismatches.
  • domain assumption Crater detector generalizes to new images if resolution and scale match training set
    Stated as a note in the abstract.

pith-pipeline@v0.9.1-grok · 5723 in / 1199 out tokens · 34968 ms · 2026-06-27T10:03:58.575163+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

23 extracted references · 11 canonical work pages

  1. [1]

    Guidance Summary and Assessment of the Chang’e-3 Powered Descent and Landing,

    S. Li, X. Jiang, and T. Tao, “Guidance Summary and Assessment of the Chang’e-3 Powered Descent and Landing,”Journal of Spacecraft and Rockets, vol.53, pp. 258-277, 2016

    2016

  2. [2]

    On Lunar On-Orbit Vision-Based Navigation: Ter- rain Mapping, Feature Tracking Driven EKF,

    L. Singh and S. Lim, “On Lunar On-Orbit Vision-Based Navigation: Ter- rain Mapping, Feature Tracking Driven EKF,”AIAA GNC Conference, 2008

    2008

  3. [3]

    Servadio, S., Parker, W., and Linares, R. (2023). Uncertainty propagation and filtering via the koopman operator in astrodynamics. Journal of Spacecraft and Rockets, 60(5), 1639-1655

    2023

  4. [4]

    ”Differential algebra-based mul- tiple gaussian particle filter for orbit determination.” Journal of Opti- mization Theory and Applications 191.2 (2021): 459-485

    Servadio, Simone, and Renato Zanetti. ”Differential algebra-based mul- tiple gaussian particle filter for orbit determination.” Journal of Opti- mization Theory and Applications 191.2 (2021): 459-485

    2021

  5. [5]

    Lunar Orbiter State Estimation Using Neural Network- Based Crater Detection,

    L. M. Downes, “Lunar Orbiter State Estimation Using Neural Network- Based Crater Detection,” S.M. thesis, Dept. of Aeronautics and Astro- nautics, Massachusetts Institute of Technology, Cambridge, MA, USA, May 2020

    2020

  6. [6]

    Deep Learning Crater Detection for Lunar Terrain Relative Navigation,

    L. M. Downes, T. J. Steiner, and J. P. How, “Deep Learning Crater Detection for Lunar Terrain Relative Navigation,”AIAA SciTech 2020 F orum, 2020

    2020

  7. [7]

    Stefano Silvestrini, Margherita Piccinin, Giovanni Zanotti, Andrea Brandonisio, Ilaria Bloise, Lorenzo Feruglio, Paolo Lunghi, Mich `ele Lavagna, Mattia Varile, Optical navigation for Lunar landing based on Convolutional Neural Network crater detector, Aerospace Sci- ence and Technology, V olume 123, 2022, 107503, ISSN 1270-9638, https://doi.org/10.1016/j...

    work page doi:10.1016/j.ast.2022.107503 2022

  8. [8]

    Robust automatic crater detection at all latitudes on Mars with Deep-learning

    Martinez Leonard, Franc ¸ois Andrieu, Fr ´ed´eric Schmidt, Hugues Talbot, Bentley Mark. Robust automatic crater detection at all latitudes on Mars with Deep-learning. Planetary and Space Science, 2025, 260, 10.1016/j.pss.2025.106053

    work page doi:10.1016/j.pss.2025.106053 2025

  9. [9]

    X. Lin, T. Li, B. Hua, L. Li and C. Zhao, ”CraterSense: Efficient and Robust Crater-Based Lunar Localization System With Structure-Aware Sparse Transformer,” in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 19, pp. 49-69, 2026, doi: 10.1109/JSTARS.2025.3633611

    work page doi:10.1109/jstars.2025.3633611 2026

  10. [10]

    Yu, R., & Xu, Z. (2025). Crater-MASN: A Multi-Scale Adaptive Se- mantic Network for Efficient Crater Detection. Remote Sensing, 17(18),

    2025

  11. [11]

    https://doi.org/10.3390/rs17183139

    work page doi:10.3390/rs17183139

  12. [12]

    N., & Pattanayak, R

    Sinha, M., Paul, S., Ghosh, M., Mohanty, S. N., & Pattanayak, R. M. (2024). Automated Lunar Crater Identification with Chandrayaan-2 TMC-2 Images using Deep Convolutional Neural Networks. Scientific reports, 14(1), 8231. https://doi.org/10.1038/s41598-024-58438-4

    work page doi:10.1038/s41598-024-58438-4 2024

  13. [13]

    Emami, T

    E. Emami, T. Ahmad, G. Bebis, A. Nefian and T. Fong, ”Crater De- tection Using Unsupervised Algorithms and Convolutional Neural Net- works,” in IEEE Transactions on Geoscience and Remote Sensing, vol. 57, no. 8, pp. 5373-5383, Aug. 2019, doi: 10.1109/TGRS.2019.2899122

    work page doi:10.1109/tgrs.2019.2899122 2019

  14. [14]

    Mimansa Sinha, Sanchita Paul, Automated detection of craters on the lunar surface using deep learning: A review with insights from Chandrayaan-2 TMC-2 data, Results in Earth Sciences, V olume 3, 2025, 100094, ISSN 2211-7148, https://doi.org/10.1016/j.rines.2025.100094

    work page doi:10.1016/j.rines.2025.100094 2025

  15. [15]

    Deep visual odometry and pose reconstruction through single image depth map and triangulation for terrain relative navigation

    Silvestrini, S. Deep visual odometry and pose reconstruction through single image depth map and triangulation for terrain relative navigation. Astrodyn (2026). https://doi.org/10.1007/s42064-025-0282-4

    work page doi:10.1007/s42064-025-0282-4 2026

  16. [16]

    Chatterjee, S

    S. Chatterjee, S. Chakraborty, P. Roy Chowdhury, B. Deshmukh and A. Nath, ”Toward Faster and Accurate Detection of Craters,” in IEEE Geoscience and Remote Sensing Letters, vol. 22, pp. 1-5, 2025, Art no. 8001605, doi: 10.1109/LGRS.2025.3557756

    work page doi:10.1109/lgrs.2025.3557756 2025

  17. [17]

    UNet++: A Nested U-Net Architecture for Medical Image Segmentation,

    Z. Zhou, M. M. R. Siddiquee, N. Tajbakhsh, and J. Liang, “UNet++: A Nested U-Net Architecture for Medical Image Segmentation,”Deep Learning in Medical Image Analysis, 2018

    2018

  18. [18]

    NASA Crater Detection Challenge,

    NASA, “NASA Crater Detection Challenge,” National Aeronautics and Space Administration, Nov. 25, 2025. [Online]. Available: https://www. nasa.gov/directorates/stmd/prizes-challenges-crowdsourcing-program/ center-of-excellence-for-collaborative-innovation-coeci/ nasa-crater-detection-challenge/. [Accessed: Mar. 1, 2026]

    2025

  19. [19]

    NASA Crater Detection,

    Topcoder, “NASA Crater Detection,” Topcoder Challenge Page. [On- line]. Available: https://www.topcoder.com/nasa-crater-detection. [Ac- cessed: Mar. 1, 2026]

    2026

  20. [20]

    Robinson, M.S., Brylow, S.M., Tschimmel, M. et al. Lunar Reconnais- sance Orbiter Camera (LROC) Instrument Overview. Space Sci Rev 150, 81–124 (2010). https://doi.org/10.1007/s11214-010-9634-2

    work page doi:10.1007/s11214-010-9634-2 2010

  21. [21]

    Robbins, S. J. (2019). A new global database of lunar impact craters ¿1–2 km: 1. Crater locations and sizes, comparisons with published databases, and global analysis. Journal of Geophysical Research: Planets, 124, 871–892. https://doi.org/10.1029/2018JE005592

    work page doi:10.1029/2018je005592 2019

  22. [22]

    Global distribution of large lunar craters: implications for resurfacing and impactor populations,

    J. W. Head, C. I. Fassett, S. J. Kadish, D. E. Smith, M. T. Zuber, G. A. Neumann, and E. Mazarico, “Global distribution of large lunar craters: implications for resurfacing and impactor populations,” Science, vol. 329, no. 5998, pp. 1504–1507, Sep 2010

    2010

  23. [23]

    Compilation of a Global GIS Crater Database for the Moon,

    N. G. Barlow et al., “Compilation of a Global GIS Crater Database for the Moon,”AAS/Division for Planetary Sciences Meeting Abstracts, 2012

    2012