SLAM 點雲圖鑑Point clouds for construction
M-01 · 2025

Kinematic-ICP

Kinematic-ICP 針對在平面上移動、配備 3D LiDAR 的輪式機器人，把單輪車（unicycle）運動學模型放進點到點 ICP 最佳化，並以輪式里程計為初值與正則化項，使估計結果符合平台運動限制。正則化強度依情況自適應調整，以在特徵不足的走廊中更信任輪式里程計。作者回報已部署於 Dexory 倉儲機器人隊伍。

里程計＋局部建圖近十年已讀全文同儕審查已出版當代 LiDAR（慣性）里程計與 SLAM

閱讀版本
arXiv v3 (2410.10277v3, 2025-02-10, 'Accepted at ICRA 2025'); IEEE ICRA 2025 version of record not opened
全文取得途徑
arXiv 預印本
本頁內容

Adds unicycle kinematic constraints and adaptive wheel-odometry regularization to point-to-point ICP for wheeled robots on planar floors, improving robustness in feature-poor corridors.

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 (Robosense Bpearl, Hesai XT32)、wheel odometry
原文測試平台wheeled UGV
狀態估計point-to-point ICP (KISS-ICP based) with unicycle kinematic model and adaptive regularization toward the wheel-odometry initial guess
資料關聯point-to-point
時間表示discrete planar poses
去畸變de-skewing in preprocessing inherited from KISS-ICP (Sec. III); motion source for de-skewing not detailed in text read
迴圈閉合none
全域最佳化none
地圖表示voxel-grid local map as in KISS-ICP (author-stated, Sec. III)
先驗資訊wheel odometry; planar-surface assumption; LiDAR-to-base extrinsic required
可輸出幾何planar odometry; export format 原文未報告
計算需求100 Hz on a single CPU core (authors' comparison with Fuse at ~10 Hz, Sec. V-C)

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
LiDARBpearl方法輸入未標示90 x 360 deg hemispherical 32-beam LiDAR, 10 Hz(Guadagnino et al., 2025b, Sec. V-A-1)
LiDARHesai LiDAR XT32方法輸入未標示10 Hz(Guadagnino et al., 2025b, Sec. V-A-1)
輪式或腿式里程計原文未報告方法輸入未標示wheel-encoder odometry of the Dexory robot and of the Husky, used as initial guess and regularization prior(Guadagnino et al., 2025b, Sec. III; Sec. V-A-1)
全測站Leica Nova MS60參考或真值量測未標示tracks a reflective prism on the robot; angular accuracy 0.0003 deg, range accuracy 3 mm or better; initially time-synchronized with the robot(Guadagnino et al., 2025b, Sec. V-A-1; Fig. 2)
載具平台Dexory robot方法輸入未標示differential drive with front and back caster wheels; extendable 12 m tower; about 500 kg; provides wheel-encoder odometry(Guadagnino et al., 2025b, Sec. V-A-1; Sec. V-B)
載具平台Clearpath Husky A200方法輸入未標示four-wheeled skid-steering robot with wheel-encoder odometry(Guadagnino et al., 2025b, Sec. V-A-1)
運算硬體原文未報告執行運算平台未標示Kinematic-ICP runs at 100 Hz on a single CPU core; Fuse at about 10 Hz(Guadagnino et al., 2025b, Sec. V-C)

作者報告的優勢與限制

優勢

  • Better RPE and ATE than wheel odometry on all seven sequences and the lowest RPE of all methods on every sequence (Table II)
  • Authors' text claims better results than KISS-ICP on all sequences and consistent outperformance of EKF and Fuse, but Table II shows exceptions in ATE: Palace 1.56 m versus 0.69 m (KISS-ICP), 0.78 m (EKF) and 0.69 m (Fuse); WO + 2D KISS-ICP lower on Campus 1 (0.26 vs 0.42 m) and Warehouse Large (4.11 vs 4.42 m) (Sec. V-C; Table II)
  • Deployed on a commercial warehouse robot fleet (abstract; Sec. VI)

限制

  • Planar-surface assumption: on uneven park terrain it performs slightly worse than KISS-ICP variants because slippage, rolling and pitching are not modelled (Sec. V-C)
  • Warehouse accuracy measured against Cartographer SLAM output rather than an independent reference (Sec. V-A)
  • Adaptive regularization is not always best: fixed beta = 0.01 gives lower Warehouse Small errors (RPE 0.39% vs 0.53%, ATE 0.20 vs 0.26 m) and larger fixed beta gives lower Palace ATE (Table III)
  • Requires the LiDAR-to-base extrinsic calibration and robot wheel odometry (Sec. III; Sec. V-A)

營建工程相關證據

在營運中倉庫（0.35 ha 至 9.45 ha）測試，屬既有建築室內；平面地板假設可能適用於樓板完成後的室內巡檢，但不適用於不平整工地地面（推論）。

原文驗證環境：已完工建築、受控實驗、獨立參考量測

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 3 個比較組，合計 19 筆紀錄。

Guadagnino et al., 2025b · Table II 本方法 14 筆

表格設定（擷取紀錄原文）：RPE is the KITTI average translation error over 1, 2, 5, 10, 20, 50 and 100 m segments (%); ATE is RMS absolute translation error after alignment (m); warehouse reference is Cartographer, campus and park reference is the total station (Guadagnino et al., 2025b, Table II)

RPE [%] (KITTI metric, 1-100 m segments)，authors' warehouse and campus sequences · Campus 0

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。

這些是 Guadagnino et al., 2025b 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。

統計量：平均值（mean）；對齊方式：原文未報告；單位：%；場景：flat campus pavement, Clearpath Husky; reference Leica Nova MS60 total station

資料來源作者報告值(Guadagnino et al., 2025b, Table II)

數值與出處
方法（原文寫法）報告值出處
Wheel Odometry4.93%(Guadagnino et al., 2025b, Table II)
KISS-ICP [31]4.9%(Guadagnino et al., 2025b, Table II)
WO + 3D KISS-ICP4.64%(Guadagnino et al., 2025b, Table II)
WO + 2D KISS-ICP4.43%(Guadagnino et al., 2025b, Table II)
EKF (robot_localization fusing WO + 2D KISS-ICP)6.28%(Guadagnino et al., 2025b, Table II)
Fuse (fixed-lag smoother fusing WO + 2D KISS-ICP)4.16%(Guadagnino et al., 2025b, Table II)
Kinematic-ICP本方法原文提出2.97%(Guadagnino et al., 2025b, Table II)

Guadagnino et al., 2025b · Table III 本方法 4 筆

表格設定（擷取紀錄原文）：Ablation on regularization of the wheel-odometry translation prior; all rows are Kinematic-ICP variants (Guadagnino et al., 2025b, Table III)

RPE [%]，authors' warehouse and campus sequences · Palace

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。

這些是 Guadagnino et al., 2025b 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。

統計量：平均值（mean）；對齊方式：原文未報告；單位：%；場景：uneven park terrain (grass, curb walks) at Poppelsdorf Palace, Bonn; reference total station

資料來源作者報告值(Guadagnino et al., 2025b, Table III)

數值與出處
方法（原文寫法）報告值出處
Fixed beta = 0.012.39%(Guadagnino et al., 2025b, Table III)
Fixed beta = 0.11.79%(Guadagnino et al., 2025b, Table III)
Fixed beta = 1.02.56%(Guadagnino et al., 2025b, Table III)
Fixed beta = 10.03.71%(Guadagnino et al., 2025b, Table III)
Fixed beta = 100.03.99%(Guadagnino et al., 2025b, Table III)
No Regularization3.72%(Guadagnino et al., 2025b, Table III)
Kinematic-ICP (adaptive regularization)本方法原文提出2.38%(Guadagnino et al., 2025b, Table III)

Guadagnino et al., 2025b · Text Sec. V-C 本方法 1 筆

指標runs at ... Hz

資料集與序列authors' sequences

表格設定（擷取紀錄原文）：Runtime comparison stated in text; Fuse value given as approximately 10 Hz (Guadagnino et al., 2025b, Text Sec. V-C)

runs at ... Hz，authors' sequences

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。

這些是 Guadagnino et al., 2025b 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。

統計量：原文未報告；對齊方式：未對齊；單位：Hz

資料來源作者報告值(Guadagnino et al., 2025b, Text Sec. V-C)

數值與出處
方法（原文寫法）報告值出處
Kinematic-ICP本方法原文提出硬體：single-core CPU (model not reported)100 Hz(Guadagnino et al., 2025b, Sec. V-C)
Fuse10 Hz(Guadagnino et al., 2025b, Sec. V-C)

來源

  • Guadagnino et al., 2025b

    Tiziano Guadagnino, Benedikt Mersch, Ignacio Vizzo, Saurabh Gupta, Meher V.R. Malladi, Luca Lobefaro, Guillaume Doisy, Cyrill Stachniss（2025）Kinematic-ICP: Enhancing LiDAR Odometry with Kinematic Constraints for Wheeled Mobile Robots Moving on Planar Surfaces2025 IEEE International Conference on Robotics and Automation (ICRA), pp. 11090-11096

    DOI 10.1109/icra55743.2025.11128503arXiv 2410.10277程式碼

    同儕審查已出版已讀全文近十年查證後修正

程式碼：https://github.com/PRBonn/kinematic-icp（授權：MIT (LICENSE file 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。

回到方法圖鑑

SLAM 點雲圖鑑

本站為綜述論文《SLAM-Based Point Clouds for Construction: A Review》研究過程的互動說明文件。互動圖為教學模擬，不代表任何特定系統的實測表現；方法描述、數值與結論以所引原始文獻為準。

內容基準日 2026-09-25 · 研究與網站製作使用 AI 輔助工具，見 證據標記與方法

選擇開啟Esc關閉