LOCUS 2.0
LOCUS 2.0 是以 LiDAR 為核心、可鬆耦合其他里程計的多階段 GICP 里程計，針對算力與記憶體受限的地下探勘機器人設計。它把 GICP 所需的點共變異數改由預先計算的法向量直接構成，地圖點不必重算共變異數；以自適應體素濾波把每幀點數維持在設定值附近，使運算時間不隨環境大小或 LiDAR 數量劇烈變動；地圖只保留以機器人為中心 50 m 的滑動視窗，可用多執行緒八元樹或 ikd-Tree 儲存，以限制記憶體用量。系統本身沒有迴圈閉合。
本頁內容
LiDAR-centric multi-stage GICP odometry for compute- and memory-constrained underground robots: point covariances built from stored normals, an adaptive voxel filter that holds the point count constant, and a 50 m sliding-window map (multi-threaded octree or ikd-tree); optional non-lidar odometry seeds the registration; no loop closure.
技術屬性
欄位內容為文獻擷取紀錄的原文用語（英文），以原文為據；「未查證」表示本研究尚未讀到該資訊，不代表該方法不具備此能力。
| 感測輸入 | one or more 3D LiDARs merged in the body frame (three Velodyne VLP16 on Husky; one lidar on Spot)、IMU for per-lidar motion distortion correction、optional non-lidar odometry (wheel-inertial, kinematic-inertial or visual-inertial) as initial guess through the sensor integration module (LiDAR-centric, loosely coupled) |
|---|---|
| 原文測試平台 | wheeled UGV (Husky)、legged (Spot) |
| 狀態估計 | multi-stage GICP registration, scan-to-scan then scan-to-submap, with point covariances built from precomputed normals (GICP from normals); health-aware loosely coupled use of optional non-lidar odometry as the scan-to-scan initial guess (Sec. III; Sec. III-A) |
| 資料關聯 | GICP nearest-neighbour correspondences (maximum correspondence distance 0.3, 20 iterations) on points reduced by an adaptive voxel grid filter that holds the point count near a set value (1000 to 10000 tested) (Sec. III-B; Sec. IV-C) |
| 時間表示 | discrete poses |
| 去畸變 | IMU-based motion distortion correction of each lidar stream in the preprocessor (Sec. III) |
| 迴圈閉合 | none (odometry system; no loop closure described) |
| 全域最佳化 | none |
| 地圖表示 | sliding-window robot-centred point map (window 50 m) stored in a multi-threaded octree (two threads alternately box-filter and rebuild) or an ikd-tree; normals are stored with map points (Sec. III-C; Sec. IV-D) |
| 先驗資訊 | none for odometry; the evaluation ground truth uses survey-grade maps |
| 可輸出幾何 | 6-DoF odometry and a local 3D point cloud map |
| 計算需求 | Husky runs 4 threads and Spot 1 thread; in Table III LOCUS 2.0 uses 61.05 to 119.00% CPU in the column printed as max (100% = one core) and 1.01 to 2.42 GB maximum memory; GICP from normals reduced the computational metrics by 18.57% on average and raised the odometry rate by 11.10% (Sec. IV-B; Sec. IV-C1; Table III) |
使用設備
原文使用的感測器、運算硬體與載具（equipment）。型號保留原文寫法，連結到設備頁中同一型號的歸併名稱；角色依原文用途分為方法輸入、資料集感測器、執行運算平台、參考或真值量測（reference or ground truth）與比較對象設備。
| 類別 | 型號（原文寫法） | 角色 | 資料集 | 原文規格 | 出處 |
|---|---|---|---|---|---|
| LiDAR | VLP16歸入：Velodyne VLP-16 | 方法輸入 | NeBula odometry dataset (DARPA SubT, Team CoSTAR) | three per Husky, extrinsically calibrated: one flat, one pitched forward 30 deg, one pitched backward 30 deg; 10 Hz (two used where marked in Table I) | (Reinke et al., 2022, Sec. IV-A; Table I) |
| LiDAR | Spot on-board lidar (model not reported) | 方法輸入 | NeBula odometry dataset (DARPA SubT, Team CoSTAR) | one lidar, extrinsically calibrated; 10 Hz | (Reinke et al., 2022, Sec. IV-A) |
| 慣性量測單元（IMU） | IMU (model not reported) | 方法輸入 | NeBula odometry dataset (DARPA SubT, Team CoSTAR) | recorded at 50 Hz; used for motion distortion correction | (Reinke et al., 2022, Sec. III; Sec. IV-A) |
| 相機 | camera streams (models not reported) | 資料集感測器 | NeBula odometry dataset (DARPA SubT, Team CoSTAR) | recorded in the dataset | (Reinke et al., 2022, Sec. IV-A) |
| 輪式或腿式里程計 | wheeled inertial odometry (WIO) on Husky | 方法輸入 | NeBula odometry dataset (DARPA SubT, Team CoSTAR) | recorded at 50 Hz | (Reinke et al., 2022, Sec. IV-A) |
| 輪式或腿式里程計 | Spot kinematic inertial odometry (KIO) and visual inertial odometry (VIO), out of the box | 資料集感測器 | NeBula odometry dataset (DARPA SubT, Team CoSTAR) | recorded in the dataset | (Reinke et al., 2022, Sec. IV-A) |
| 載具平台 | Husky | 方法輸入 | NeBula odometry dataset (DARPA SubT, Team CoSTAR) | skid-steer wheeled robot on rough terrain | (Reinke et al., 2022, Sec. IV-A; Fig. 4) |
| 載具平台 | Spot | 方法輸入 | NeBula odometry dataset (DARPA SubT, Team CoSTAR) | legged robot | (Reinke et al., 2022, Sec. IV-A; Fig. 4) |
| 其他 | survey-grade 3D map (provided by DARPA or produced by the team; instrument not reported) | 參考或真值量測 | NeBula odometry dataset (DARPA SubT, Team CoSTAR) | ground-truth trajectory produced by LOCUS 1.0 scan-to-survey-map registration with manual post-processing | (Reinke et al., 2022, Sec. IV-A) |
作者報告的優勢與限制
優勢
- Smallest max and mean APE on 4 of the 6 compared underground datasets (A, C, F and H); on I LINS and on J FAST-LIO have lower APE, although the text states 5 of 6 (Table III; Sec. IV-E)
- Only compared method that did not fail in the Bruceton Mine tunnel dataset F (Table III; Sec. IV-E)
- Adaptive voxel filter keeps the point count and callback time consistent across environments and lidar configurations (Sec. IV-C2; Figs. 6-8)
- Sliding-window maps cut memory by 38.88% to 87.76% relative to the LOCUS 1.0 octree baseline (Table II)
- Open-source code and an 11 h, 16 km underground dataset with Husky and Spot data (abstract; Sec. IV-A)
限制
- GICP from normals raises mean and max APE by 10.82% on average (5.23% without tunnel dataset F) because normals come from sparse clouds and are not recomputed (Sec. IV-C1)
- Sliding-window maps increase CPU use by 9.36% to 50.42% relative to the static octree (Table II)
- ikd-tree insertion and search take on average 222% and 140% more time than the octree (Sec. IV-D1)
- Smaller sliding maps bound memory but give higher APE (Sec. IV-D2)
- Memory use is higher than FAST-LIO and LINS on most datasets (Table III; Sec. IV-E)
- Ground truth is LOCUS 1.0 run against a survey-grade map with manual post-processing, not an independent trajectory measurement (Sec. IV-A)
營建工程相關證據
未在施工現場測試；資料來自 DARPA SubT 的停用電廠、礦坑隧道、地鐵站與熔岩洞，其中電廠與地鐵站屬既有建築或基礎設施。狹長走廊與隧道的幾何退化和地下工程施工相似，作者報告在隧道資料 F 中 FAST-LIO 與 LINS 誤差極大而 LOCUS 2.0 仍可運作；但真值是以 LOCUS 1.0 對測量等級地圖配準產生，並非完全獨立的量測。另一篇基礎設施檢測回顧的作者表示無法執行 LOCUS 的程式碼（Ghadimzadeh Alamdari et al., 2025）。
原文驗證環境：地下或隧道、基礎設施、已完工建築
報告的性能數據
以下是原文作者報告的性能數值（author-reported results），不是本研究重新量測的結果。每張圖只並列同一個比較組（comparison group，同一張表、同一組實驗設定）內的方法；不同比較組之間的數值不可直接比較，也不構成排名。
本方法共出現在 4 個比較組，合計 44 筆紀錄。
Reinke et al., 2022 · Table III 本方法 30 筆
表格設定（擷取紀錄原文）：Underground datasets A, C, F, H, I, J (Table I); LOCUS 2.0 versus FAST-LIO and LINS; column labels reproduced as printed (APE max [m], APE mean [%], CPU [%] max and mean, max memory [GB]); many printed 'max' values are below 'mean' values; ground truth from LOCUS 1.0 against survey-grade maps (Reinke et al., 2022, Table III)
APE max [m]，NeBula odometry dataset (DARPA SubT, Team CoSTAR) · A: power plant, Elma WA (urban), Husky, 631.53 m
只並列這張表在相同設定下報告的方法；以「本方法：」開頭者為本頁方法。失敗、未執行與未報告以標記呈現，不是 0。
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 Reinke et al., 2022 在此表設定下報告的數值（author-reported results），只能在同一個比較組內對照，不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Reinke et al., 2022, Table III)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LOCUS 2.0本方法原文提出 | 0.19 m | (Reinke et al., 2022, Table III) |
| FAST-LIO | 0.79 m | (Reinke et al., 2022, Table III) |
| LINS | 0.43 m | (Reinke et al., 2022, Table III) |
Reinke et al., 2022 · Table II 本方法 8 筆
資料集與序列NeBula odometry dataset (DARPA SubT, Team CoSTAR) · datasets used in Sec. IV-D (F and I shown in Fig. 9)
表格設定（擷取紀錄原文）：Relative memory and CPU change of sliding-window map structures versus the LOCUS 1.0 static octree with 0.001 m leaf (baseline), 50 m map window, GICP from normals (Reinke et al., 2022, Table II)
Memory (relative change)，NeBula odometry dataset (DARPA SubT, Team CoSTAR) · datasets used in Sec. IV-D (F and I shown in Fig. 9)
只並列這張表在相同設定下報告的方法；以「本方法：」開頭者為本頁方法。失敗、未執行與未報告以標記呈現，不是 0。
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 Reinke et al., 2022 在此表設定下報告的數值（author-reported results），只能在同一個比較組內對照，不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Reinke et al., 2022, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LOCUS 2.0 with ikd-tree本方法原文提出 | -68.09 % change | (Reinke et al., 2022, Table II) |
| LOCUS 2.0 with mto 0.001 (multi-threaded octree, leaf 0.001 m)本方法原文提出 | -38.88 % change | (Reinke et al., 2022, Table II) |
| LOCUS 2.0 with mto 0.01本方法原文提出 | -62.15 % change | (Reinke et al., 2022, Table II) |
| LOCUS 2.0 with mto 0.1本方法原文提出 | -87.76 % change | (Reinke et al., 2022, Table II) |
Reinke et al., 2022 · Text Sec. IV-C1 本方法 4 筆
資料集與序列NeBula odometry dataset (DARPA SubT, Team CoSTAR) · A-J (average)
表格設定（擷取紀錄原文）：GICP from normals versus standard GICP inside LOCUS 2.0, averaged over datasets A-J and 5 runs each; percentage changes stated in text (Reinke et al., 2022, Text Sec. IV-C1)
average reduction of computational metrics (CPU, delay, registration and callback times)，NeBula odometry dataset (DARPA SubT, Team CoSTAR) · A-J (average)
這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Reinke et al., 2022 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LOCUS 2.0, GICP from normals vs GICP本方法原文提出 | -18.57 % change | (Reinke et al., 2022, Sec. IV-C1) |
Reinke et al., 2022 · Text Sec. IV-D1 本方法 2 筆
資料集與序列NeBula odometry dataset (DARPA SubT, Team CoSTAR) · all datasets (average)
表格設定（擷取紀錄原文）：Map structure operation timing relative to the octree (Reinke et al., 2022, Text Sec. IV-D1)
insertion computation time relative to octree，NeBula odometry dataset (DARPA SubT, Team CoSTAR) · all datasets (average)
這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Reinke et al., 2022 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LOCUS 2.0 with ikd-tree本方法原文提出 | 222 % more | (Reinke et al., 2022, Sec. IV-D1) |
來源
Reinke et al., 2022
（2022）LOCUS 2.0: Robust and Computationally Efficient Lidar Odometry for Real-Time 3D MappingIEEE Robotics and Automation Letters, 7(4):9043-9050
DOI 10.1109/lra.2022.3181357arXiv 2205.11784程式碼
同儕審查已出版已讀全文近十年查證後修正
相關版本
- 預印本：LOCUS 2.0 (arXiv v2, accepted RA-L version; arXiv title adds 'Underground') https://arxiv.org/abs/2205.11784
- 前身方法：LOCUS: A Multi-Sensor Lidar-Centric Solution for High-Precision Odometry and 3D Mapping in Real-Time (RA-L 2021; not read) https://doi.org/10.1109/LRA.2020.3044864
- 資料集：NeBula odometry dataset released with the paper https://github.com/NeBula-Autonomy/nebula-odometry-dataset
- 程式碼釋出：NeBula-Autonomy/LOCUS https://github.com/NeBula-Autonomy/LOCUS
程式碼：https://github.com/NeBula-Autonomy/LOCUS（授權：MIT (LICENSE file read)）。有公開程式碼不等於已被重現，也不代表目前版本與論文版本相同。