LION
LION 是 CoSTAR 團隊參加 DARPA 地下挑戰賽所用的 LiDAR 慣性里程計。前端以廣義 ICP 做相鄰掃描配準並先以 IMU 旋轉對齊重力，後端在 GTSAM 中以 3 秒固定延遲滑動視窗平滑器融合 IMU 預積分與掃描間相對位姿，同時線上估計 LiDAR 與 IMU 外參；作者明確說明它是鬆耦合架構，且不建地圖、不做迴圈閉合。另以點到平面 ICP Hessian 平移區塊的條件數作為可觀性指標，條件數過大時通知監督邏輯 HeRO 改用其他里程計來源，例如輪式慣性里程計。
本頁內容
Loosely coupled fixed-lag smoother (GTSAM iSAM2, 3 s window) that fuses IMU preintegration with scan-to-scan GICP relative poses, estimates LiDAR-IMU extrinsics online, and reports the condition number of the translational point-to-plane ICP Hessian as an observability score so a supervisor can switch odometry sources in degenerate tunnels.
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 (model not named)、IMU (model not named) |
|---|---|
| 原文測試平台 | wheeled UGV (team CoSTAR ground robots in mines and offices)、simulation (extrinsic calibration test) |
| 狀態估計 | fixed-lag sliding-window smoother (3 s window) in GTSAM solved with iSAM2, fusing IMU preintegration factors with relative-pose factors from LiDAR odometry; loosely coupled because points or scans are not in the state (Sec. 2, 3.1, 4) |
| 資料關聯 | scan-to-scan Generalized ICP between consecutive clouds, each pre-rotated into a gravity-aligned frame with the IMU rotation as initial guess; no feature extraction; LOCUS can replace the front-end (Sec. 2, 4) |
| 時間表示 | discrete poses in a sliding window with IMU preintegration (Sec. 2) |
| 去畸變 | 原文未報告 |
| 迴圈閉合 | none within LION; drift is compensated by the separate LAMP mapping system (Sec. 4) |
| 全域最佳化 | none within LION |
| 地圖表示 | none (LION builds no map) (Sec. 4) |
| 先驗資訊 | none; LiDAR-IMU extrinsics (rotation and translation) estimated online in the state (Sec. 2) |
| 可輸出幾何 | gravity-aligned odometry at up to 200 Hz plus an observability (condition-number) score for a supervisory switching logic (HeRO) (Sec. 2, 3.1) |
| 計算需求 | back-end tuned to use about 30% of one CPU core of an Intel i7 NUC; LiDAR odometry at 10 Hz, output up to 200 Hz (Sec. 3.1) |
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 | 3D LiDAR (model not stated) | 方法輸入 | 未標示 | LiDAR odometry computed at 10 Hz | (Tagliabue et al., 2021, Sec. 2; Sec. 3.1) |
| 慣性量測單元（IMU） | IMU (model not stated) | 方法輸入 | 未標示 | IMU and LION output provided at up to 200 Hz | (Tagliabue et al., 2021, Sec. 2; Sec. 3.1) |
| 輪式或腿式里程計 | wheel odometry (encoders, model not stated) | 比較對象設備 | 未標示 | fused with the IMU in an EKF for the Wheel-Inertial baseline and as HeRO fallback | (Tagliabue et al., 2021, Sec. 3.1; Sec. 3.2) |
| 載具平台 | team CoSTAR ground robots (not further specified) | 方法輸入 | 未標示 | explored the Arch Coal Mine about 275 m underground | (Tagliabue et al., 2021, Fig. 1; Sec. 1) |
| 運算硬體 | Intel NUC with i7 processor (model not stated) | 執行運算平台 | 未標示 | LION back-end used about 30% of one CPU core | (Tagliabue et al., 2021, Sec. 3.1) |
作者報告的優勢與限制
優勢
- Fusing IMU with scan-to-scan odometry cut position drift strongly, e.g. Track A Run 2 from 18.72 m (4.11%) to 7.00 m (1.53%) (Table 1)
- Position RMSE comparable to LOAM on the shorter runs (Track A Run 2 7.00 vs 7.22 m; Track B Run 2 3.78 vs 5.55 m), with gravity-aligned output at IMU rate rather than LiDAR rate; the authors report comparable or slightly lower roll and pitch errors in Fig. 5 (Sec. 3.1, Table 1)
- Online extrinsic calibration recovered a 0.1 m simulated offset within about 20 s (Sec. 3.1, Fig. 7)
- Condition number rose above 13 in a featureless corridor and above 10 along a mine shaft, and switching to wheel-inertial odometry reduced a return error from about 9 m to about 1 m (Sec. 3.2, Figs. 8-10)
- Uses about 30% of one NUC i7 core (Sec. 3.1)
限制
- No map and no loop closure; the odometry frame drifts slowly and relies on LAMP for correction (Sec. 4)
- LOAM achieved lower position and yaw drift on longer trajectories thanks to its map, e.g. 10.99 m versus 56.92 m on Track A Run 1 (Sec. 3.1, Table 1)
- Attitude RMSE in Table 1 is higher than LOAM on all four runs (0.36, 0.10, 0.27, 0.05 rad versus 0.14, 0.08, 0.21, 0.03 rad) (Table 1)
- Loosely coupled design chosen for shared compute and modularity rather than accuracy (Sec. 4)
- Reference trajectories come from the team's own LAMP SLAM system, not an independent survey (Sec. 3.1)
- Observability threshold is user defined (Sec. 2)
營建工程相關證據
論文在美國 NIOSH 實驗礦坑、煤礦與金礦等地下環境以及 JPL 辦公室走廊測試，未涉及營建工地，參考軌跡為團隊自己的 LAMP 輸出而非獨立測量。其以 ICP Hessian 平移部分條件數偵測長走廊與隧道退化，並交由監督邏輯切換到輪式慣性里程計的做法，可直接對應施工中隧道與長廊的定位風險管理（推論）。
原文驗證環境：地下或隧道、已完工建築、模擬
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 14 筆紀錄。
Tagliabue et al., 2021 · Table 1 本方法 12 筆
表格設定（擷取紀錄原文）：DARPA SubT Tunnel Circuit (NIOSH experimental mines, Pittsburgh), one robot; LAMP output used as ground truth; LiDAR odometry 10 Hz; LION sliding window 3 s (Tagliabue et al., 2021, Table 1)
t(m) position RMSE，DARPA SubT Tunnel Circuit runs · Track A Run 1 (685 m, 1520 s)
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 Tagliabue et al., 2021 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(Tagliabue et al., 2021, Table 1)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Wheel-Inertial | 130.5 m | (Tagliabue et al., 2021, Table 1) |
| Scan-To-Scan | 105.47 m | (Tagliabue et al., 2021, Table 1) |
| LION本方法原文提出 | 56.92 m | (Tagliabue et al., 2021, Table 1) |
| LOAM | 10.99 m | (Tagliabue et al., 2021, Table 1) |
Tagliabue et al., 2021 · Text Sec.3.2 本方法 2 筆
指標total error when the robot goes back to the original position (approximately)
資料集與序列JPL office-like environment · corridor loop
表格設定（擷取紀錄原文）：Office-like environment with a featureless corridor section; total translation error when the robot returns to the start, before loop closure; values given as approximately (Tagliabue et al., 2021, Text Sec.3.2)
total error when the robot goes back to the original position (approximately)，JPL office-like environment · corridor loop
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 Tagliabue et al., 2021 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(Tagliabue et al., 2021, Text Sec.3.2)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LION without observability module本方法 | 9 m | (Tagliabue et al., 2021, Sec. 3.2; Fig. 10) |
| LION with observability module (HeRO switches to WIO in the corridor)本方法原文提出 | 1 m | (Tagliabue et al., 2021, Sec. 3.2; Fig. 10) |
來源
Tagliabue et al., 2021
（2021）LION: Lidar-Inertial Observability-Aware Navigator for Vision-Denied EnvironmentsExperimental Robotics (ISER 2020), Springer Proceedings in Advanced Robotics 19, SPAR 19, pp. 380-390
DOI 10.1007/978-3-030-71151-1_34arXiv 2102.03443
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 2102.03443 v1 (2021-02-05), only version https://arxiv.org/abs/2102.03443