MOLA-LO
MOLA-LO 主張以「視圖式地圖」（view-based map：帶時間戳的原始感測資料加上位姿）作為基本地圖表示，事後可依任務重新產生各種度量地圖，例如點雲、雜湊體素、佔據體素或類 NDT 地圖。建圖流程可像組合神經網路層一樣以可重用區塊設定，而不需撰寫程式。其 LiDAR 里程計在類 ICP 最佳化中緊耦合估計線速度與角速度，不需 IMU；迴圈閉合則以後處理方式進行，並可加入 GNSS 做地理參考。
本頁內容
An open framework centred on view-based maps from which task-specific metric maps are regenerated, with IMU-free LiDAR odometry that jointly estimates velocities inside ICP, offline loop closure, and GNSS georeferencing.
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 | 3D LiDAR (16 to 128 rings)、2D LiDAR、optional wheel odometry for kinematic prediction、optional consumer-grade GNSS (loop closure and georeferencing only)、IMU not used by LO |
|---|---|
| 原文測試平台 | vehicle (KITTI, KITTI-360, MulRan, ParisLuco, UAL campus)、handheld (Newer College)、UAV (Voxgraph, HILTI 2021 drone arena, NTU VIRAL)、legged (ANYmal C, DARPA SubT)、backpack (Almeria forests)、wheelchair and indoor robot with 2D LiDAR |
| 狀態估計 | ICP-like optimizer with tightly-coupled estimation of linear and angular velocity (no IMU required); self-adaptive parameters via dynamic variables |
| 資料關聯 | Default lidar3d-default configuration: point-to-point pairings between the sparser twice-decimated scan layer and the local map, solved by Gauss-Newton on SE(3) with a robust kernel; matching threshold and kernel scale follow an adaptive threshold inspired by KISS-ICP but driven by a proportional feedback controller on ICP quality; the 3D-NDT configuration adds point-to-plane pairings for planar voxels |
| 時間表示 | discrete scan poses with linear and angular velocities estimated inside the ICP-like optimizer and used for intra-scan SE(3) interpolation (Sec. 3.3.1; Sec. 7.1) |
| 去畸變 | per-scan de-skewing by trajectory interpolation on SE(3) using the estimated velocities (Eq. 2); ablation on UAL campus data: ATE 7.97 m without vs 6.65 m with de-skewing (Sec. 7.1) |
| 迴圈閉合 | Post-processing only (not concurrent with LO): the view-based map is split into sub-maps with bounding boxes and optional GNSS georeferencing; candidates are sub-map pairs whose expected bounding-box intersection (Monte Carlo over relative poses from Dijkstra on the sub-map graph) exceeds a threshold, with no place-recognition descriptor; each candidate is verified by an ICP pipeline with separate ground and non-ground layers and a voxel-occupancy quality score |
| 全域最佳化 | Two-level graph: key-frame factor graph in GTSAM with LO relative-pose factors and optional GNSS factors, optimized first without and then with robust kernels, re-optimized after each accepted loop closure; the sub-map graph is used only for candidate search |
| 地圖表示 | View-based map (key-frames with pose, velocities and raw observations) as the stored map; the LO local map is a single hashed-voxel point cloud with at most 20 points per voxel and resolution 1.5% of the estimated maximum sensor range clamped to 0.5 to 1.0 m, updated only when a decider's distance criteria are met; other layers (contiguous point clouds, VDB occupancy voxels, 3D-NDT, 2D grids) can be regenerated from the view-based map |
| 先驗資訊 | optional GNSS; prior metric map for localization |
| 可輸出幾何 | arbitrary metric maps regenerated from view-based maps; georeferenced maps and trajectories exportable to KML (Sec. 5.2) |
| 計算需求 | No GPU use is reported; per-scan times measured on an Intel i7-8700 at 3.20 GHz for MulRan (24 ms default, 39 ms 3D-NDT, 40 to 42 ms with loop closure); other tables report 23 ms (KITTI), 13.1 ms (DARPA SubT), 14.5 ms (UAL campus) and 31.6 to 69.4 ms (NTU VIRAL configurations); results are deterministic across CPUs |
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 | OS0-128歸入：Ouster OS0-128 | 資料集感測器 | Newer College extension (Zhang et al. 2021) | 128 rings; handheld (Table 1; Sec. 4.9 text names it OS1-128, inconsistent with Table 1) | (Blanco-Claraco, 2025, Table 1; Sec. 4.9) |
| LiDAR | OS1-64歸入：Ouster OS1-64 | 資料集感測器 | Newer College (Ramezani et al. 2020) | 64 rings; handheld (Table 1; Sec. 4.9 text names it OS0-64, inconsistent with Table 1) | (Blanco-Claraco, 2025, Table 1; Sec. 4.9) |
| LiDAR | OS1-64歸入：Ouster OS1-64 | 資料集感測器 | MulRan | vehicle; dataset also provides consumer-grade GNSS used only in loop closure | (Blanco-Claraco, 2025, Table 1; Sec. 4.1) |
| LiDAR | Velodyne HDL-64E | 資料集感測器 | KITTI odometry | 64 rings; 0.205 deg vertical angle correction applied | (Blanco-Claraco, 2025, Table 1; Sec. 4.2) |
| LiDAR | HDL-64E歸入：Velodyne HDL-64E | 資料集感測器 | KITTI-360 | vehicle | (Blanco-Claraco, 2025, Table 1; Sec. 4.3) |
| LiDAR | OS1-64歸入：Ouster OS1-64 | 資料集感測器 | Voxgraph | drone; RTK-based ground truth | (Blanco-Claraco, 2025, Table 1; Sec. 4.7) |
| LiDAR | OS0-64歸入：Ouster OS0-64 | 資料集感測器 | HILTI 2021 | drone testing arena, industrial unit | (Blanco-Claraco, 2025, Table 1; Sec. 4.6) |
| LiDAR | Velodyne HDL-32歸入：Velodyne HDL-32E | 資料集感測器 | ParisLuco | 32 rings; vehicle | (Blanco-Claraco, 2025, Table 1; Sec. 4.4) |
| LiDAR | OS0-32歸入：Ouster OS0-32 | 方法輸入 | Almeria forests (Aguilar et al. 2024) | backpack kit with the LiDAR as the only sensor; forests | (Blanco-Claraco, 2025, Table 1; Sec. 5.3) |
| LiDAR | 2 x OS1-16 (Table 1); text names two Velodyne VLP-16 | 資料集感測器 | NTU-VIRAL | one horizontal and one vertical LiDAR on a drone; motion capture ground truth | (Blanco-Claraco, 2025, Table 1; Sec. 4.5) |
| LiDAR | Velodyne VLP-16 | 資料集感測器 | DARPA Subterranean final event | one per ANYmal C robot | (Blanco-Claraco, 2025, Table 1; Sec. 4.8) |
| LiDAR | Velodyne VLP-16 | 資料集感測器 | UAL VLP-16 campus | electric vehicle on university campus | (Blanco-Claraco, 2025, Table 1; Sec. 4.10) |
| LiDAR | SICK LMS 2D range finder | 資料集感測器 | Freiburg building 079 (fr079) | 2D LiDAR; Pioneer2 robot in text, PowerBot in Table 1 | (Blanco-Claraco, 2025, Table 1; Sec. 5.1) |
| LiDAR | SICK LMS 2D range finder | 資料集感測器 | Malaga CS faculty | 2D LiDAR on a robotic wheelchair with encoders, 1.9 km | (Blanco-Claraco, 2025, Table 1; Sec. 5.1) |
| 地面雷射掃描儀（TLS） | survey-quality scanners (models not reported) | 參考或真值量測 | DARPA Subterranean final event | ground truth point cloud used to derive ground-truth trajectories by scan matching | (Blanco-Claraco, 2025, Sec. 4.8) |
| GNSS 接收器 | consumer-grade GNSS receiver | 方法輸入 | MulRan | used for georeferencing and loop-closure candidate search, not by LO | (Blanco-Claraco, 2025, Sec. 4.1; Sec. 3.11; Sec. 7.5) |
| GNSS 接收器 | RTK GNSS | 參考或真值量測 | UAL VLP-16 campus | 3D positioning used as ground truth | (Blanco-Claraco, 2025, Sec. 4.10) |
| 輪式或腿式里程計 | wheel encoders | 方法輸入 | fr079 and Malaga CS faculty | incremental odometry used by kinematic state prediction in the 2D configuration | (Blanco-Claraco, 2025, Sec. 3.8; Sec. 5.1) |
| 載具平台 | ANYmal C legged robot | 資料集感測器 | DARPA Subterranean final event | four robots, one sequence each | (Blanco-Claraco, 2025, Sec. 4.8) |
| 運算硬體 | Intel i7-8700 at 3.20 GHz | 執行運算平台 | 未標示 | desktop CPU used for MulRan per-scan timings | (Blanco-Claraco, 2025, Sec. 4.1) |
| 其他 | motion capture system | 參考或真值量測 | NTU-VIRAL | accurate ground truth for NTU VIRAL | (Blanco-Claraco, 2025, Sec. 4.5) |
作者報告的優勢與限制
優勢
- Single self-adaptive configuration tested on 83 sequences over more than 250 km of automotive, handheld, airborne and quadruped data (abstract)
- Enables posterior generation of maps optimized for different tasks (abstract; Sec. 3.2)
限制
- Fails to converge on some Hilti 2021 sequences without public ground truth (narrow, featureless indoor spaces) where feature extraction or multi-modality seem needed (Sec. 4.6)
- Vertical (z) drift dominates error in some ANYmal SubT sequences (Sec. 4.8)
- Loop closure runs as post-processing, not concurrently (Sec. 3.12)
- Loop closure and georeferencing are not released as open source; the other framework components are (Sec. 8)
- With a single horizontal LiDAR on the NTU VIRAL drone, ATE is about an order of magnitude worse than LIO baselines; only a two-LiDAR near and far map configuration reaches LIO-level errors (Sec. 4.5; Table 6)
營建工程相關證據
在 Hilti 2021 的 RPG Drone Testing Arena 序列（作者表列為無人機、工業廠房、約 0.08 km）評估；作者並回報在部分狹窄、無特徵且未公開地面真值的 Hilti 2021 室內序列無法收斂。另有背包式森林建圖與 DARPA SubT 洞穴（ANYmal）案例；未見施工現場點雲幾何評估。
原文驗證環境：公開基準、地下或隧道
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 9 個比較組，合計 92 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 5 組列在最後，並連到性能比較頁。
Blanco-Claraco, 2025 · Table 9 本方法 42 筆
指標absolute translational error (RMSE, evo_ape -a)
表格設定（擷取紀錄原文）：Handheld Newer College sequences; ATE RMSE from evo_ape -a (Umeyama alignment); x(value) marks divergence; no method uses the IMU; same default configuration for all datasets (Blanco-Claraco, 2025, Table 9)
absolute translational error (RMSE, evo_ape -a)，Newer College (2020, sequences 01 and 02) · 01
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。
- 發散
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 Blanco-Claraco, 2025 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(Blanco-Claraco, 2025, Table 9)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| KISS-ICP | 0.61 m | (Blanco-Claraco, 2025, Table 9) |
| MOLA-LO (ours)本方法原文提出 | 0.68 m | (Blanco-Claraco, 2025, Table 9) |
| MOLA-LO (always updates local map)本方法 | 5.11 m發散註記（擷取紀錄）：diverged (value printed in parentheses) | (Blanco-Claraco, 2025, Table 9) |
| MOLA-LO + LC (ours)本方法原文提出 | 0.31 m | (Blanco-Claraco, 2025, Table 9) |
Blanco-Claraco, 2025 · Table 4 本方法 16 筆
指標absolute translational error (RMSE, evo_ape -a)
表格設定（擷取紀錄原文）：KITTI-360 ATE RMSE (evo_ape -a); sequences 03, 07 and 10 have no loop closures (Blanco-Claraco, 2025, Table 4)
absolute translational error (RMSE, evo_ape -a)，KITTI-360 · 00
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 Blanco-Claraco, 2025 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(Blanco-Claraco, 2025, Table 4)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| KISS-ICP | 5.5 m | (Blanco-Claraco, 2025, Table 4) |
| MOLA-LO (ours)本方法原文提出 | 2.81 m | (Blanco-Claraco, 2025, Table 4) |
| MOLA-LO + LC (ours)本方法原文提出 | 0.72 m | (Blanco-Claraco, 2025, Table 4) |
Blanco-Claraco, 2025 · Table 3 本方法 12 筆
資料集與序列KITTI odometry · Avr. 00-10
表格設定（擷取紀錄原文）：KITTI odometry average RTE over training sequences 00 to 10; IMLS-SLAM, MULLS and CT-ICP2 values from their publications; 0.205 deg vertical correction applied to KITTI scans for KISS-ICP, SiMpLE and MOLA-LO (Blanco-Claraco, 2025, Table 3)
Avr. 00-10 RTE (%)，KITTI odometry · Avr. 00-10
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 Blanco-Claraco, 2025 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(Blanco-Claraco, 2025, Table 3)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| IMLS-SLAM | 0.55% | (Blanco-Claraco, 2025, Table 3) |
| KISS-ICP | 0.55% | (Blanco-Claraco, 2025, Table 3) |
| SiMpLE (offline) | 0.55% | (Blanco-Claraco, 2025, Table 3) |
| SiMpLE (online) | 0.62% | (Blanco-Claraco, 2025, Table 3) |
| MOLA-LO (default) (ours)本方法原文提出 | 0.55% | (Blanco-Claraco, 2025, Table 3) |
| MOLA-LO (3D-NDT) (ours)本方法原文提出 | 0.58% | (Blanco-Claraco, 2025, Table 3) |
| MOLA-LO (Horn's) (ours)本方法 | 0.65% | (Blanco-Claraco, 2025, Table 3) |
| MULLS (LC) | 0.52% | (Blanco-Claraco, 2025, Table 3) |
| CT-ICP2 (LC) | 0.53% | (Blanco-Claraco, 2025, Table 3) |
| MOLA-LO (default) + LC (ours)本方法原文提出 | 0.58% | (Blanco-Claraco, 2025, Table 3) |
Blanco-Claraco, 2025 · Table 10 本方法 6 筆
資料集與序列UAL VLP-16 campus dataset · UAL campus
表格設定（擷取紀錄原文）：UAL campus, electric vehicle with VLP-16, RTK GNSS ground truth; with and without scan deskewing (Blanco-Claraco, 2025, Table 10)
absolute translational error (RMSE, evo_ape -a)，UAL VLP-16 campus dataset · UAL campus
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 Blanco-Claraco, 2025 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(Blanco-Claraco, 2025, Table 10)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| KISS-ICP (w/o deskew) | 13.21 m | (Blanco-Claraco, 2025, Table 10) |
| MOLA-LO (w/o deskew) (ours)本方法 | 7.97 m | (Blanco-Claraco, 2025, Table 10) |
| KISS-ICP | 10.06 m | (Blanco-Claraco, 2025, Table 10) |
| MOLA-LO (ours)本方法原文提出 | 6.65 m | (Blanco-Claraco, 2025, Table 10) |
| MOLA-LO + LC (ours)本方法原文提出 | 1 m | (Blanco-Claraco, 2025, Table 10) |
其他比較組
來源
Blanco-Claraco, 2025
（2025）A flexible framework for accurate LiDAR odometry, map manipulation, and localizationThe International Journal of Robotics Research, 44(9):1553-1599
DOI 10.1177/02783649251316881arXiv 2407.20465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：A flexible framework for accurate LiDAR odometry, map manipulation, and localization (arXiv v3) https://arxiv.org/abs/2407.20465
- 程式碼釋出：MOLAorg/mola and MOLAorg/mola_lidar_odometry https://github.com/MOLAorg/mola
程式碼：https://github.com/MOLAorg/mola（授權：mola_lidar_odometry: GPL-3.0 (LICENSE and package.xml checked); main mola repository: per-package licenses declared in package.xml (not individually checked)）。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。