LAMP 2.0
LAMP 2.0 是 CoSTAR 團隊為 DARPA 地下挑戰賽開發的集中式多機器人 LiDAR 位姿圖 SLAM。各機器人的前端介面可接不同里程計（LOCUS 或 Hovermap）與不同 LiDAR 配置，先以 HeRO 狀態估計去除掃描畸變、合併多顆 LiDAR，再用自適應體素濾波讓點數一致，並每約 2 m 或 30 度建立關鍵節點與對應掃描送到基地站。基地站的多機器人前端以自適應半徑產生迴圈閉合候選，依可觀測性、圖神經網路預測效益與 RSSI 排序，再以 TEASER++ 或 SAC-IA 初始對齊後用 GICP 精修；後端以 GNC 搭配 Levenberg-Marquardt 在 GTSAM 中做抗離群值的位姿圖最佳化。作者在煤礦、核電廠、SubT 決賽場地與石灰岩礦的四組資料上評估，並釋出含地面真值的資料集。
本頁內容
Centralized multi-robot LiDAR pose-graph SLAM for large underground sites: odometry-agnostic front-end interfaces, prioritized inter- and intra-robot loop closures computed with TEASER++ or SAC-IA plus GICP, and a GNC outlier-robust back-end in GTSAM; evaluated on four datasets from a coal mine, a power plant, the SubT Finals course and a limestone mine with released ground truth.
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 LiDARs (three Velodyne lidars on Husky, a single lidar on Spot; models not reported)、Hovermap payload on some robots、odometry input from LOCUS or Hovermap (front-end agnostic) |
|---|---|
| 原文測試平台 | wheeled UGV (Husky)、legged (Spot)、multi-robot teams of up to four robots with a centralized base station |
| 狀態估計 | Centralized multi-robot pose-graph optimization in GTSAM with Levenberg-Marquardt and Graduated Non-Convexity (GNC) for outlier-robust inlier selection, optionally after Incremental Consistency Maximization (ICM); single-robot front-ends send sparse pose graphs (key nodes every about 2 m or 30 deg) with keyed scans (Sec. II-B, II-D) |
| 資料關聯 | Proximity-based loop-closure candidates with adaptive radius, prioritized by observability (ICP information-matrix eigenvalues), a GNN-predicted graph benefit and RSSI beacons; relative pose from TEASER++ or SAC-IA initialization refined by GICP, rejecting poor alignments (Sec. II-C) |
| 時間表示 | discrete key nodes; scans de-skewed by the HeRO local state estimate (Sec. II-B) |
| 去畸變 | motion distortion corrected with the Heterogeneous Robust Odometry (HeRO) local state estimate before merging multi-lidar scans (Sec. II-B) |
| 迴圈閉合 | intra- and inter-robot loop closures from the multi-robot front-end with two-stage registration (TEASER++ or SAC-IA, then GICP) (Sec. II-C) |
| 全域最佳化 | centralized multi-robot pose-graph optimization with GNC (and ICM) outlier rejection in GTSAM (Sec. II-D) |
| 地圖表示 | pose graph with keyed scans (adaptive voxel-filtered point clouds); optimized global point-cloud map formed by transforming keyed scans with optimized poses (Sec. II-D) |
| 先驗資訊 | common reference frame from a gate with three reflective plates of known coordinates at the entrance (Sec. II-A) |
| 可輸出幾何 | globally consistent multi-robot trajectories and point-cloud map; map compared with surveyed ground-truth map by cloud-to-cloud error (Fig. 4) |
| 計算需求 | base station during SubT: AMD Ryzen Threadripper 3990x (64 cores); paper experiments: laptop Intel i7-8750H (12 cores), data played back in real time (Sec. III-A, III-D) |
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 | Velodyne lidars (three per Husky; model not reported) | 方法輸入 | CoSTAR multi-robot datasets | merged after extrinsic calibration and adaptive voxelization | (Chang et al., 2022, Sec. II-B, III-A) |
| LiDAR | single lidar on Spot (model and manufacturer not reported) | 方法輸入 | CoSTAR multi-robot datasets | one of the two Spot sensor configurations (the other is a Hovermap); point clouds differ in size and density from the Husky three-lidar setup | (Chang et al., 2022, Sec. III-A) |
| 行動掃描設備 | Hovermap | 方法輸入 | CoSTAR multi-robot datasets | provides odometry and point clouds as an alternative front-end | (Chang et al., 2022, Sec. II-B, III-A) |
| 載具平台 | Husky | 方法輸入 | CoSTAR multi-robot datasets | wheeled platform equipped with three Velodyne lidars and a Hovermap | (Chang et al., 2022, Sec. III-A) |
| 載具平台 | Spot | 方法輸入 | CoSTAR multi-robot datasets | quadruped platform equipped with either a single lidar or a Hovermap | (Chang et al., 2022, Sec. III-A) |
| 運算硬體 | AMD Ryzen Threadripper 3990x (64 cores) | 執行運算平台 | 未標示 | portable base-station workstation during the SubT Challenge | (Chang et al., 2022, Sec. III-A) |
| 運算硬體 | laptop with Intel i7-8750H (12 cores) | 執行運算平台 | 未標示 | runs the experiments reported in the paper | (Chang et al., 2022, Sec. III-A) |
| 其他 | three reflective plates on an entrance gate | 方法輸入 | 未標示 | fiducial markers with known 3D coordinates for initial pose calibration | (Chang et al., 2022, Sec. II-A) |
作者報告的優勢與限制
優勢
- Per-robot ATE below 2 m for trajectories up to 2.2 km (Table IV; Sec. III-D)
- TEASER++ or SAC-IA initialization lowered false-positive rates and loop-closure pose errors compared with odometric initialization (Table II)
- Fewer candidates but more verified and inlier loop closures than LAMP 1.0; GNC handled more than 80% outlier loop closures (Table III; Sec. III-C)
- Handles heterogeneous robots with different lidar configurations and odometry sources in the same run (Sec. II-B, III-A)
- Map errors below 4 m against surveyed maps in all four large environments (Fig. 4; Sec. III-D)
限制
- Centralized architecture may not scale to large robot teams; a distributed version is future work (Sec. IV)
- Loop-closure recall in the wide KU limestone tunnels stays low across initializations (11.3% to 29.0%, the highest with GT initialization) (Table II)
- Initial common frame relies on a calibration gate with reflective markers (Sec. II-A)
- Map accuracy is shown only as colour-coded cloud-to-cloud error maps (Fig. 4), not tabulated
營建工程相關證據
驗證場景是地下礦坑、廢棄核電廠與人工洞穴，屬地下基礎設施而非施工工地；釋出的位姿圖、關鍵掃描與以測量地圖產生的真值，可作為隧道與地下工程 SLAM 的大尺度參考資料（Sec. I、III-B）。
原文驗證環境：地下或隧道、獨立參考量測、跨場域
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 2 個比較組，合計 38 筆紀錄。
Chang et al., 2022 · Table IV 本方法 26 筆
指標ATE [m] (text: average trajectory error)
表格設定（擷取紀錄原文）：End-to-end system evaluation: data played back in real time to the base station (about 1 h per run), same odometry input for all variants; ground-truth trajectories from scan-to-map localization in surveyed global maps; per-robot ATE (Chang et al., 2022, Table IV)
ATE [m] (text: average trajectory error)，CoSTAR multi-robot dataset: Tunnel · husky3 (traversed 1194 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 Chang 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(Chang et al., 2022, Table IV)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LAMP 2.0本方法原文提出 | 0.65 m | (Chang et al., 2022, Table IV) |
| LAMP 2.0 single robot (no inter-robot loop closures)本方法原文提出 | 0.83 m | (Chang et al., 2022, Table IV) |
| LAMP 1.0 | 1.07 m | (Chang et al., 2022, Table IV) |
Chang et al., 2022 · Table III 本方法 12 筆
表格設定（擷取紀錄原文）：Number of loop closures at each stage of the multi-robot front-end and back-end; LAMP 1.0 uses fixed-radius candidates, odometric ICP initialization and ICM; LAMP 2.0 adds adaptive radius, prioritization, SAC-IA or TEASER++ initialization and GNC (Chang et al., 2022, Table III)
# Generated (loop-closure candidates)，CoSTAR multi-robot dataset: Tunnel · Tunnel
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 Chang 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(Chang et al., 2022, Table III)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LAMP 1.0 | 22206 count | (Chang et al., 2022, Table III) |
| LAMP 2.0本方法原文提出 | 9755 count | (Chang et al., 2022, Table III) |
來源
Chang et al., 2022
（2022）LAMP 2.0: A Robust Multi-Robot SLAM System for Operation in Challenging Large-Scale Underground EnvironmentsIEEE Robotics and Automation Letters, 7(4):9175-9182
DOI 10.1109/lra.2022.3191204arXiv 2205.13135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：arXiv 2205.13135 v1 to v3 https://arxiv.org/abs/2205.13135
- predecessor：LAMP: Large-scale autonomous mapping and positioning for exploration of perceptually-degraded subterranean environments, ICRA 2020, pp. 80-86 (ref. [8]; DOI not verified) not_verified
- 程式碼釋出：NeBula-Autonomy/LAMP https://github.com/NeBula-Autonomy/LAMP
- dataset release：NeBula-Autonomy/nebula-multirobot-dataset (pose graphs, keyed scans, ground truth) https://github.com/NeBula-Autonomy/nebula-multirobot-dataset
程式碼：https://github.com/NeBula-Autonomy/LAMP（授權：MIT (LICENSE file on main branch checked 2026-09-25)）。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。