Wildcat
Wildcat 是 CSIRO 的線上 3D LiDAR 慣性 SLAM，其里程計是 Zebedee 等離線連續時間方法概念的即時實作：在固定長度的滑動時間視窗內，把點雲依位置與時間聚成多解析度橢球面元（surfel），以面元對面元的點到面型成本與 IMU 成本共同修正取樣位姿，再以三次 B-spline 內插得到高頻軌跡並重投影面元，藉此處理運動畸變。後端以六秒子地圖為節點做位姿圖最佳化，加入重力方向項並合併重疊節點，使計算量隨探索空間而非任務時間成長，亦支援多機去中心化建圖。
本頁內容
Wildcat makes Zebedee-style continuous-time surfel odometry online in a sliding window with IMU fusion and B-spline interpolation, and adds submap-based pose-graph optimization with node merging and multi-agent support.
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 Velodyne VLP-16 in two configurations: servo-spun at 0.5 Hz on an inclined mount for 120 deg vertical FoV with the measurement rate set to 20 Hz (SpinningPack), or fixed with the native 30 deg vertical FoV (FlatPack; rate not stated); Ouster OS1-64 at 10 Hz with 120 m range in MulRan DCC03、IMU (9-DoF 3DM-CV5 at 100 Hz in the SpinningPack; FlatPack IMU model not stated) |
|---|---|
| 原文測試平台 | handheld、legged (Boston Dynamics Spot)、tracked robot、vehicle |
| 狀態估計 | sliding-window continuous-time trajectory optimization (Gauss-Newton with Cauchy IRLS) over sample-pose corrections, followed by pose-graph optimization with a gravity-direction term (Cauchy IRLS) (Sec. IV-V) |
| 資料關聯 | multi-resolution surfels (voxel clustering by position and time, planarity-filtered ellipsoids); reciprocal kNN matching in a 7-D descriptor space; point-to-plane-type surfel cost weighted by lidar noise and surfel thickness (Sec. IV-A to IV-C) |
| 時間表示 | continuous-time: cubic B-spline interpolation between corrected sample poses, linear interpolation in cost terms (Sec. IV-B, IV-C) |
| 去畸變 | surfels re-projected with the continuously updated IMU-rate trajectory (Sec. IV-B2) |
| 迴圈閉合 | candidates by Mahalanobis-distance search or place recognition (e.g., Scan Context); point-to-plane ICP between surfel submaps, with global initialization when uncertainty is large; Mahalanobis gating (Sec. V-B2) |
| 全域最佳化 | pose graph over submap nodes with node merging for overlapping submaps; supports decentralized multi-agent mapping (Sec. V) |
| 地圖表示 | local multi-resolution surfel maps bundled into six-second submaps (Sec. V-A) |
| 先驗資訊 | none |
| 可輸出幾何 | globally optimized multi-agent point cloud map (e.g., submitted to DARPA SubT) and trajectory (Sec. VI-B) |
| 計算需求 | odometry about 63.3 ms average (about 15 Hz) on an Intel Xeon W-10885M laptop CPU; about 1 to 4 Hz on the pack's NVIDIA Jetson AGX Xavier; odometry memory plateaus at about 500 MiB and PGO memory stays below 3 GiB on QCAT SpinningPack (Sec. VI-E, Figs. 15, 18) |
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 VLP-16 | 方法輸入 | DARPA SubT Final Event (Team CSIRO Data61); QCAT (SpinningPack) | measurement rate 20 Hz; mounted at an inclined angle on a servomotor spinning about the sensor z axis at 0.5 Hz, giving 120 deg vertical FoV (SpinningPack) | (Ramezani et al., 2022, Sec. VI-A1, VI-A3; Table I) |
| LiDAR | Velodyne VLP-16 | 方法輸入 | QCAT (FlatPack) | fixed, vertical FoV 30 deg (FlatPack) | (Ramezani et al., 2022, Sec. VI-A3; Table I) |
| LiDAR | Ouster OS1-64 | 資料集感測器 | MulRan (DCC03) | 10 Hz, range 120 m, vehicle-mounted | (Ramezani et al., 2022, Sec. VI-A2; Table I) |
| 地面雷射掃描儀（TLS） | survey-grade laser scanner (model not stated) | 參考或真值量測 | DARPA SubT Final Event | DARPA ground-truth point cloud used at 1 cm resolution; about 100 person-hours according to DARPA | (Ramezani et al., 2022, Sec. VI-B) |
| 行動掃描設備 | SpinningPack (CSIRO perception pack) | 方法輸入 | DARPA SubT Final Event; QCAT (SpinningPack) | spinning VLP-16, 3DM-CV5 IMU, four RGB cameras; robot-mounted at DARPA and hand-held at QCAT | (Ramezani et al., 2022, Sec. VI-A1, VI-A3; Fig. 6) |
| 行動掃描設備 | FlatPack (CSIRO hand-held perception pack) | 方法輸入 | QCAT (FlatPack) | fixed VLP-16 with 30 deg vertical FoV | (Ramezani et al., 2022, Sec. VI-A3; Fig. 6) |
| 慣性量測單元（IMU） | 3DM-CV5 | 方法輸入 | DARPA SubT Final Event; QCAT (SpinningPack) | 9-DoF, angular velocity and linear acceleration at 100 Hz (SpinningPack); FlatPack IMU model not stated | (Ramezani et al., 2022, Sec. VI-A1) |
| GNSS 接收器 | GPS (model not stated) | 參考或真值量測 | MulRan (DCC03) | combined with fiber optic gyro and SLAM to give 6-DoF ground truth at 100 Hz | (Ramezani et al., 2022, Sec. VI-A2) |
| 相機 | RGB camera (four per SpinningPack, model not stated) | 資料集感測器 | DARPA SubT Final Event; QCAT (SpinningPack) | part of the SpinningPack; not used in the described Wildcat pipeline | (Ramezani et al., 2022, Sec. VI-A1) |
| 載具平台 | Boston Dynamics Spot (two robots) | 方法輸入 | DARPA SubT Final Event | legged robots carrying SpinningPacks | (Ramezani et al., 2022, Sec. VI-A1) |
| 載具平台 | BIA5 ATR tracked robot (two robots) | 方法輸入 | DARPA SubT Final Event | tracked robots carrying SpinningPacks | (Ramezani et al., 2022, Sec. VI-A1) |
| 載具平台 | vehicle (MulRan data collection car, model not stated) | 資料集感測器 | MulRan (DCC03) | urban driving in Daejeon, sequence about 5 km | (Ramezani et al., 2022, Sec. VI-A2) |
| 運算硬體 | Intel Xeon W-10885M | 執行運算平台 | 未標示 | laptop CPU | (Ramezani et al., 2022, Sec. VI-E) |
| 運算硬體 | NVIDIA Jetson AGX Xavier | 執行運算平台 | 未標示 | perception pack onboard computer; odometry about 1 to 4 Hz | (Ramezani et al., 2022, Sec. VI-E) |
| 其他 | fiber optic gyro (model not stated) | 參考或真值量測 | MulRan (DCC03) | part of the MulRan ground-truth solution | (Ramezani et al., 2022, Sec. VI-A2) |
| 其他 | surveyed targets (63; survey instrument not stated) | 參考或真值量測 | QCAT | scattered over indoor, outdoor, 3-storey office and mock-up tunnel areas | (Ramezani et al., 2022, Sec. VI-A3, VI-D; Fig. 12) |
作者報告的優勢與限制
優勢
- DARPA SubT Final Event: point-wise comparison of the 40 cm-voxelised multi-agent map to the 1 cm surveyed map gave 3 cm mean and 5 cm std error, with more than 95% of points within 10 cm after fine alignment (Sec. VI-B, Fig. 8)
- QCAT handheld sequences (~5 km each, 63 surveyed targets): mean absolute target error 0.42 m (FlatPack) and 0.34 m (SpinningPack), lower than LIO-SAM and FAST-LIO2 on the targets those systems reached before failing (Sec. VI-D, Table II)
- Field tested on handheld, legged, tracked and car platforms (Sec. I, VII)
- With loop closure enabled on MulRan DCC03, APE was slightly lower than LIO-SAM (Sec. VI-C; values shown only as box plots in Fig. 9)
- One common voxel-filter parameter set for all datasets, whereas LIO-SAM and FAST-LIO2 were tuned per scenario for QCAT (Sec. VI-D)
- Node merging reduced 1402 submaps to 213 pose-graph nodes (about 85 %) on QCAT SpinningPack, with PGO memory below 3 GiB (Sec. VI-E, Figs. 16, 18)
限制
- Onboard Jetson AGX Xavier odometry runs at only about 1-4 Hz (Sec. VI-E)
- Future work needed on odometry resilience across more environments and on loop-closure detection (Sec. VII)
- Technology is subject of a PCT patent application and no official code is released (Acknowledgement; no repository in paper)
- On MulRan DCC03 Wildcat odometry drift (2.9 %, 0.01 deg/m) was slightly higher than LIO-SAM odometry (2.4 %, 0.009 deg/m) (Sec. VI-C)
- QCAT accuracy relies on manually picked target centres and MSAC alignment, and the baselines are scored on smaller target subsets than Wildcat, so the Table II comparison is not on identical targets (Sec. VI-D, Table II)
營建工程相關證據
論文引言提及 LiDAR 慣性 SLAM 可作為營建等測量的低成本替代方案，但僅為動機陳述；實驗包含地下礦坑式賽道（以測量級雷射掃描地圖為參考）與含室內、隧道、樓梯的 QCAT 園區（以 60 個以上測量標靶為參考），未在營建工地驗證。
原文驗證環境：地下或隧道、公開基準、獨立參考量測、跨場域、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 8 個比較組，合計 36 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 4 組列在最後，並連到性能比較頁。
Ramezani et al., 2022 · Table II 本方法 8 筆
表格設定（擷取紀錄原文）：QCAT surveyed targets (63 in total): point-to-point error between manually picked target centres in each map and surveyed positions after MSAC robust alignment; RMSE computed during alignment; LIO-SAM and FAST-LIO2 only on the subset of targets identified before they failed (mainly in the tunnel); baselines had tuned voxel-filter parameters, Wildcat used one common setting (Ramezani et al., 2022, Table II)
absolute error mean，QCAT (in-house) · FlatPack
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。
- 不完整
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 Ramezani 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(Ramezani et al., 2022, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LIO-SAM [3] | 0.92 m不完整註記（擷取紀錄）：partial: failed before completing the sequence; statistics cover only identified targets | (Ramezani et al., 2022, Table II) |
| FAST-LIO2 [7] | 1.09 m不完整註記（擷取紀錄）：partial: failed before completing the sequence; statistics cover only identified targets | (Ramezani et al., 2022, Table II) |
| Wildcat (ours)本方法原文提出 | 0.42 m | (Ramezani et al., 2022, Table II) |
Zhang et al., 2023c · Fig. 9 (printed table) 本方法 8 筆
指標RMSE ATE [cm]
表格設定（擷取紀錄原文，這些數值分屬表中不同部分）：(Zhang et al., 2023c, Fig. 9 (printed table))
- Per-sequence RMSE ATE (cm) of the top three teams, printed as numbers inside Fig. 9; bold marks sub-cm results
- See first Fig. 9 row
RMSE ATE [cm]，Hilti-Oxford (Hilti SLAM Challenge 2022) · Exp11 Lower Gallery
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 Zhang et al., 2023c 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(Zhang et al., 2023c, Fig. 9 (printed table))
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| CSIRO (Wildcat SLAM)本方法 | 0.9 cm | (Zhang et al., 2023c, Fig. 9) |
| Vision&Robotics (MC2SLAM) | 0.7 cm | (Zhang et al., 2023c, Fig. 9) |
| HKU (FastLIO2, BALM) | 0.9 cm | (Zhang et al., 2023c, Fig. 9) |
Ramezani et al., 2022 · Text Sec.VI-E 本方法 7 筆
資料集與序列QCAT (in-house) · SpinningPack
表格設定（擷取紀錄原文）：Runtime and memory of Wildcat running online on QCAT SpinningPack (Ramezani et al., 2022, Text Sec.VI-E)
average runtime of the main odometry optimisation loop，QCAT (in-house) · SpinningPack
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 Ramezani 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 |
|---|---|---|
| Wildcat odometry本方法原文提出硬體：laptop with Intel Xeon W-10885M CPU | 63.3 ms | (Ramezani et al., 2022, Sec. VI-E, Fig. 15) |
Ramezani et al., 2022 · Text Sec.VI-B 本方法 5 筆
資料集與序列DARPA SubT Final Event · prize run (multi-agent)
表格設定（擷取紀錄原文）：DARPA SubT Final Event prize run, four robots with SpinningPacks: Wildcat map voxelised at 40 cm compared point-wise with the 1 cm DARPA survey map after a fine alignment (method not specified); DARPA's own scoring quoted separately (Ramezani et al., 2022, Text Sec.VI-B)
average distance error between corresponding points，DARPA SubT Final Event · prize run (multi-agent)
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 Ramezani 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 |
|---|---|---|
| Wildcat multi-agent map本方法原文提出 | 0.03 m | (Ramezani et al., 2022, Sec. VI-B, Fig. 7) |
其他比較組
來源
Ramezani et al., 2022
（2022）Wildcat: Online Continuous-Time 3D Lidar-Inertial SLAMarXiv, arXiv:2205.12595
預印本已讀全文近十年查證後修正
相關版本
- 預印本：arXiv 2205.12595 v1 (IEEE T-RO template header, no journal version found on 2026-09-25) https://arxiv.org/abs/2205.12595