Construction and Verification of SpaceMat-LiDAR Dataset for Spacecraft Material Classification in On-Orbit Servicing 面向在軌服務太空器材料分類的 SpaceMat-LiDAR 資料集建構與驗證
- 針對航天材料識別難、LiDAR 資料匱乏問題，於模擬太空環境下採集並建立了 SpaceMat-LiDAR 真實資料集。該資料集涵蓋 MLI、CFRP 等六類典型航天材料，系統覆蓋了不同掃描距離與入射角，填補了領域空白，並確立了規範的採集標註流程。
- 針對航天材料幾何同質與強反射挑戰，建構融合幾何與強度之深度學習模型，引入物理先驗分支利用距離／角度解耦強度偏移，有效識別 MLI 與 CFRP 等材料。實驗驗證其性能顯著優於單一特徵基線，為「天工計畫」在軌作業提供了可靠感知支撐。
- 資料集價值需透過基線實驗量化驗證。實作並對比多種材料分類基線，評估強度輔助與幾何先驗對消解材料歧義的貢獻，形成公開可比的 benchmark 口徑；SpaceMat-LiDAR 填補航天材料真實 LiDAR 資源空白，為自主在軌作業中的材料感知系統提供實證基礎。
摘要
As on-orbit servicing (OOS) missions transition toward greater autonomy, the precise identification of target materials has become a critical prerequisite for complex operations, including non-cooperative pose estimation, autonomous grasping, and obstacle avoidance. However, spacecraft components frequently exhibit “geometric homogeneity,” where distinct structural elements share identical or similar geometries, rendering shape-based recognition insufficient. Furthermore, the complex optical properties of space coatings—characterized by high specularity and unique backscattering profiles—pose significant challenges for traditional passive sensors. Despite these requirements, the field lacks real-world LiDAR datasets dedicated to space-grade materials.
In this paper, we present SpaceMat-LiDAR, a pioneering real-world aerospace material dataset captured using a solid-state LiDAR in a controlled ground-based darkroom to simulate space illumination conditions. The dataset encompasses six representative categories: Multi-Layer Insulation (MLI), Carbon Fiber Reinforced Polymer (CFRP), matte black paint, solar panels, bare aluminum, and white-painted aluminum. To capture the multi-dimensional scattering characteristics of these materials, data were systematically collected across a diverse range of scanning distances and incidence angles.
Our methodology focuses on the fusion of 3D spatial coordinates and reflection intensity features. Furthermore, we investigate the effects of incorporating geometric priors—specifically distance and incidence angles—as auxiliary inputs to the classification models. Several baseline algorithms are evaluated to benchmark the contribution of intensity-aided and geometric features in resolving material ambiguity. The SpaceMat-LiDAR dataset fills a critical gap in existing data resources and provides an empirical foundation for developing material-aware perception systems for future autonomous on-orbit operations.