A Reproducible Reconstruction and Evaluation Framework for Photorealistic Lunar Analog Scene Reconstruction Toward Digital Twins

International Conference on Space Robotics (iSpaRo) 2026

Antar Mazumder1,*   Ryan Hartzell1   Luke Boyd1   Alex Robinson1   Neha Sharma2   Chance Carnival1   Frances Zhu1

* Corresponding author
1Dept. of Mechanical Engineering, Colorado School of Mines    2Dept. of Electrical Engineering, Colorado School of Mines

Paper Code Dataset Cite
Rover POV traversing the reconstructed Digital Twin — the best-performing splat reconstruction, converted to USDZ and driven live in Isaac Sim 5.

Abstract

In this paper, we present a reconstruction pipeline that empirically quantifies the photorealistic scene reconstruction accuracy from both human-controlled and rover mast camera monocular feeds against physical ground truth as a function of Lunar traversal effects (e.g., motion blur, haze due to lofted regolith, trajectory bias, sparse views). First, we benchmark six variants of 3D Gaussian Splatting (3DGS) across three Lunar analog scenes on held-out test sets using standard photometric metrics. We then establish how under traversal effects, this test-set-specific Novel View Synthesis (NVS) performance does not proportionally scale to the same photometric metrics in the reconstructed DTs.

Our results quantify that splat training metrics alone are insufficient predictors of Cross-Reconstruction Photorealistic NVS Fidelity (CRPNVS), and that factors such as rover trajectory bias and geometric scene coverage play a critical role in determining the photorealistic quality of the reconstructed scenes targeting Lunar analog Digital Twins. Additionally, we pilot study Coverage Score (Sc), a geometry-based diagnostic derived from COLMAP poses and point cloud structure that provides a meaningful heuristic about the highest achievable CRPNVS structural fidelity, independently of splat training performance in the absence of a human-controlled reference.

These findings establish a standardized, human control-referenced evaluation pipeline for photorealistic scene reconstruction for Lunar-analog Digital Twin development and future planetary exploration.

Human-control GoPro reconstruction vs rover RealSense D456 reconstruction, showing higher held-out SSIM for the rover splat despite visibly worse reconstruction quality
Why held-out metrics mislead — the rover-derived splat scores a higher held-out SSIMTr (0.92) than the human-controlled one (0.82), despite reconstructing the scene far less faithfully. Trajectory bias inflates training metrics without reflecting true novel-view fidelity.

What the Paper Establishes

Method

End-to-end pipeline: hand-held and rover captures, duplicate removal, LightGlue and COLMAP, 3DGS variant training, fiducial-marker alignment, CRPNVS assessment, and USDZ digital twin asset generation for Isaac Sim 5
Capture-to-DT pipeline — hand-held (N-shaped sweep) and rover (forward-linear) captures are each reconstructed via COLMAP + 3DGS variants, aligned on shared fiducial markers, cross-assessed for CRPNVS, then exported as a simulation-ready USDZ digital twin asset.
Human-controlled capture — the N-shaped vertical + lateral sweep used to scan each Lunar analog scene for the clean, human-controlled reference reconstruction.

Results

Ground truth, human-controlled, and rover-derived reconstructions compared side by side across three Lunar analog scenes
GT vs. RCtrl vs. RRov across all three scenes — rover-derived reconstructions show visibly more haze and structural loss than the human-controlled splats, which closely track ground truth.
COLMAP camera trajectory scatter plots for Scene 2 and Scene 3, comparing GoPro and rover coverage metrics
Coverage explains the gap — Scene 3's less linear rover traverse yields higher Wf and Wd than Scene 2 despite harsher dust, explaining its superior CRPNVS SSIM.

Simulation-Ready Digital Twins

Physics and sensors on the reconstructed DTs — rigid bodies dropped onto the scanned terrain of Scenes 1–3, a wipe between the Gaussian splat and its collision mesh (Scenes 2 and 3), and simulated Ouster OS0 lidar returns (green), in Isaac Sim 6.
ROS 2 sensor simulation — a Nova Carter driving through the Scene 1 digital twin in Isaac Sim (left) while its simulated 3D lidar point cloud streams to RViz over ROS 2 (right).
Application: wheel–terrain interaction — a rover driving on simulated granular terrain, with the reconstructed digital twin as the photorealistic backdrop.

Resources

BibTeX

@inproceedings{mazumder2026lunarcrpnvs,
      author    = {Mazumder, Antar and Hartzell, Ryan and Boyd, Luke and
                   Robinson, Alex and Sharma, Neha and Carnival, Chance and
                   Zhu, Frances},
      title     = {A Reproducible Reconstruction and Evaluation Framework for
                   Photorealistic Lunar Analog Scene Reconstruction Toward
                   Digital Twins},
      booktitle = {International Conference on Space Robotics (iSpaRo)},
      year      = {2026}
}