RESPLE
RESPLE 把三次 B 樣條(B-spline)直接嵌入狀態空間模型,以遞迴式(濾波)方式估計六自由度連續時間運動,而非以滑動視窗最佳化擬合樣條。狀態向量由位置控制點與姿態控制點增量組成,以修改後的迭代擴展卡爾曼濾波更新,每個 LiDAR 點以其時間戳在樣條上求位姿後計算點到平面殘差。同一骨幹可組成 LiDAR-only、LiDAR-慣性、多 LiDAR 與多 LiDAR-慣性里程計。
本頁內容
A recursive (filter-based) continuous-time estimator that embeds 6-DoF cubic B-splines in the state and updates spline control points with a modified iterated EKF, supporting single or multiple LiDARs with or without IMU.
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 | one or multiple 3D LiDARs (Ouster OS1-16, Livox Mid70, Mid360, Avia, Hesai XT32)、IMU optional (VN100 or LiDAR built-in IMU) |
|---|---|
| 原文測試平台 | UAV (NTU VIRAL)、vehicle (MCD)、legged (GrandTour ANYmal D)、wearable helmet (HelmDyn)、wheeled bipedal robot DIABLO (R-Campus, about 1400 m; end-to-end error reported) |
| 狀態估計 | recursive Bayesian estimator: modified iterated EKF over cubic B-spline control points (position control points and orientation increments), without error-state formulation |
| 資料關聯 | point-to-plane residual per point, plane fitted from N=5 neighbours in ikd-Tree |
| 時間表示 | continuous-time cubic B-spline (knot frequency 100 Hz in experiments) |
| 去畸變 | not required: each point evaluated at its own timestamp on the spline |
| 迴圈閉合 | none (backend for global correction listed as future work, Sec. VI) |
| 全域最佳化 | none |
| 地圖表示 | point map in ikd-Tree |
| 先驗資訊 | none |
| 可輸出幾何 | continuous-time trajectory and point map; export format 原文未報告 |
| 計算需求 | Laptop Intel i7-11800H, 48 GB RAM, Ubuntu 22.04; RESPLE node 0.97 to 4.46 ms per 10 ms observation batch (Table V); on HD_03 1.40 ms (LO) and 1.74 ms (LIO) with runtime efficiency 0.14 and 0.17 versus 0.23 (Traj-LO), 0.79 (CTE-MLO) and 3.31 (SLICT2) using 5 CPU threads (Table VI) |
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 | Ouster OS1-16 (horizontal)歸入:Ouster OS1-16 | 資料集感測器 | NTU VIRAL | drone; adopted LiDAR for RESPLE | (Cao et al., 2025, Table I; Sec. V-B1) |
| LiDAR | Livox Mid70歸入:Livox MID70 | 資料集感測器 | MCD | large-scale urban, fast ground vehicle | (Cao et al., 2025, Table I) |
| LiDAR | Hesai XT32 (L1) | 資料集感測器 | GrandTour | on the Boxi multi-sensor rig | (Cao et al., 2025, Table I; Table III) |
| LiDAR | Livox Mid360 (L2)歸入:Livox MID-360 | 資料集感測器 | GrandTour | second LiDAR; its built-in IMU used as I | (Cao et al., 2025, Table I; Table III) |
| LiDAR | Livox Mid360歸入:Livox MID-360 | 方法輸入 | HelmDyn (own experiment) | helmet-mounted, 12 x 12 x 8 m3 space, walking, running, jumping and in-hand waving | (Cao et al., 2025, Table I; Sec. V-C; Fig. 4A) |
| LiDAR | Livox Avia | 方法輸入 | R-Campus (own experiment) | on the wheeled bipedal robot | (Cao et al., 2025, Sec. V-C; Fig. 4B) |
| 慣性量測單元(IMU) | VN100 | 資料集感測器 | NTU VIRAL | adopted IMU | (Cao et al., 2025, Table I) |
| 慣性量測單元(IMU) | VN100 | 資料集感測器 | MCD | adopted IMU | (Cao et al., 2025, Table I) |
| 慣性量測單元(IMU) | built-in IMU of Livox Mid360 | 資料集感測器 | GrandTour | used as I in GrandTour | (Cao et al., 2025, Table I) |
| 慣性量測單元(IMU) | built-in IMU of Livox Mid360 | 方法輸入 | HelmDyn (own experiment) | used for R-LIO on HelmDyn | (Cao et al., 2025, Table I) |
| 載具平台 | ANYmal D quadruped robot with Boxi rig | 資料集感測器 | GrandTour | 71 Swiss environments, 15 km over 8 hours in the dataset | (Cao et al., 2025, Sec. V-B3) |
| 載具平台 | DIABLO wheeled bipedal robot | 方法輸入 | R-Campus (own experiment) | about 1400 m campus route at 1.2 m/s | (Cao et al., 2025, Sec. V-C; Fig. 4B) |
| 運算硬體 | laptop with Intel i7-11800H CPU, 48 GB RAM | 執行運算平台 | 未標示 | Ubuntu 22.04; all evaluations | (Cao et al., 2025, Sec. V) |
| 其他 | Qualisys motion capture: 12 Oqus 700+ and 8 Arqus A12 cameras with passive markers | 參考或真值量測 | HelmDyn (own experiment) | submillimeter, low latency | (Cao et al., 2025, Sec. V-C) |
作者報告的優勢與限制
優勢
- Comparable or superior accuracy and robustness to Traj-LO, CTE-MLO and FAST-LIO2 across NTU VIRAL, MCD, GrandTour and HelmDyn (abstract; Sec. V)
- LiDAR-only variant had only one failure on GrandTour sequences where compared systems failed multiple times (Sec. V-B)
- Lowest APE on all 10 HelmDyn sequences against Traj-LO, CTE-MLO, FAST-LIO2, Point-LIO and SLICT2 (Table IV)
- Fastest in the runtime comparison, 2x to 9x faster than real time (Tables V and VI)
限制
- No global correction backend; potential degeneracy to be addressed with visual sensors in future work (Sec. VI)
- Point-wise (batch size 1) or very large observation batches reduce robustness through false associations, especially at low knot frequency (Sec. V-E)
- Two underground GrandTour sequences (JTL, JTS) and PKH are not in the public release, so that evidence cannot be reproduced (Table III)
營建工程相關證據
GrandTour 評估序列中有兩段地下序列(JTL、JTS),但作者註明未公開釋出;其餘為都市、森林與山區。另有頭盔式 HelmDyn 室內高動態資料與校園雙輪足機器人序列。未見施工現場資料或點雲幾何精度評估。
原文驗證環境:公開基準、地下或隧道、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 4 個比較組,合計 74 筆紀錄。
Cao et al., 2025 · Table II (NTU VIRAL) 本方法 36 筆
指標APE (RMSE, meters)
表格設定(擷取紀錄原文):NTU VIRAL drone sequences, horizontal OS1-16 for RESPLE; APE RMSE with the official NTU VIRAL evaluation script; T-LO, C-MLO (two LiDARs) and F-LIO2 values are copied from refs. [13] (Traj-LO), [16] (CTE-MLO) and [18] (Nguyen et al. 2024, not the FAST-LIO2 paper), respectively, per the Table II footnote; x marks failure (Cao et al., 2025, Table II (NTU VIRAL))
APE (RMSE, meters),NTU VIRAL · eee_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。
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 Cao et al., 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(Cao et al., 2025, Table II (NTU VIRAL))
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| T-LO (Traj-LO) | 0.055 m | (Cao et al., 2025, Table II) |
| C-MLO (CTE-MLO, 2 LiDARs) | 0.08 m | (Cao et al., 2025, Table II) |
| F-LIO2 (FAST-LIO2) | 0.069 m | (Cao et al., 2025, Table II) |
| R-LO (RESPLE LiDAR-only)本方法原文提出 | 0.044 m | (Cao et al., 2025, Table II) |
| R-LIO (RESPLE LiDAR-inertial)本方法原文提出 | 0.036 m | (Cao et al., 2025, Table II) |
Cao et al., 2025 · Table III (GrandTour) 本方法 32 筆
指標APE (RMSE, meters)
表格設定(擷取紀錄原文):GrandTour ANYmal D quadruped with Boxi rig: L1 Hesai XT32, L2 Livox Mid360, I built-in IMU of L2; APE RMSE via evo after interpolating estimates at ground-truth timestamps; x marks failure; sequences marked * are not in the public release (Cao et al., 2025, Table III (GrandTour))
APE (RMSE, meters),GrandTour · JTL
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。
- 失敗
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 Cao et al., 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(Cao et al., 2025, Table III (GrandTour))
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| T-LO (L1) | 0.046 m | (Cao et al., 2025, Table III) |
| C-MLO (L1+L2) | 無數值失敗註記(擷取紀錄):failed | (Cao et al., 2025, Table III) |
| F-LIO2 (L1+I) | 無數值失敗註記(擷取紀錄):failed | (Cao et al., 2025, Table III) |
| R-LO (L1)本方法原文提出 | 0.035 m | (Cao et al., 2025, Table III) |
| R-MLO (L1+L2)本方法原文提出 | 0.026 m | (Cao et al., 2025, Table III) |
| R-LIO (L1+I)本方法原文提出 | 0.028 m | (Cao et al., 2025, Table III) |
| R-MLIO (L1+L2+I)本方法原文提出 | 0.028 m | (Cao et al., 2025, Table III) |
Cao et al., 2025 · Table VI 本方法 4 筆
資料集與序列HelmDyn (own dataset) · HD_03
表格設定(擷取紀錄原文):Runtime comparison on HD_03 (helmet Livox Mid360); processing time per available interval (available time 50 ms for T-LO and SLICT2, 10 ms for the others) (Cao et al., 2025, Table VI)
Processing time (ms),HelmDyn (own dataset) · HD_03
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 Cao et al., 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(Cao et al., 2025, Table VI)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| T-LO (Traj-LO)硬體:laptop, Intel i7-11800H CPU, 48 GB RAM, Ubuntu 22.04; 5 CPU threads for all systems in the runtime comparison | 11.55 ms | (Cao et al., 2025, Table VI) |
| C-MLO (CTE-MLO)硬體:laptop, Intel i7-11800H CPU, 48 GB RAM, Ubuntu 22.04; 5 CPU threads for all systems in the runtime comparison | 7.85 ms | (Cao et al., 2025, Table VI) |
| SLICT2硬體:laptop, Intel i7-11800H CPU, 48 GB RAM, Ubuntu 22.04; 5 CPU threads for all systems in the runtime comparison | 165.73 ms | (Cao et al., 2025, Table VI) |
| R-LO本方法原文提出硬體:laptop, Intel i7-11800H CPU, 48 GB RAM, Ubuntu 22.04; 5 CPU threads for all systems in the runtime comparison | 1.4 ms | (Cao et al., 2025, Table VI) |
| R-LIO本方法原文提出硬體:laptop, Intel i7-11800H CPU, 48 GB RAM, Ubuntu 22.04; 5 CPU threads for all systems in the runtime comparison | 1.74 ms | (Cao et al., 2025, Table VI) |
Cao et al., 2025 · Text Sec. V-C (R-Campus) 本方法 2 筆
指標end-to-end error
資料集與序列R-Campus (own experiment) · R-Campus
表格設定(擷取紀錄原文):Own R-Campus sequence: Livox Avia on the DIABLO wheeled bipedal robot, about 1400 m at 1.2 m/s, start and end at the same place; end-to-end error (Cao et al., 2025, Text Sec. V-C (R-Campus))
end-to-end error,R-Campus (own experiment) · R-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 Cao et al., 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(Cao et al., 2025, Text Sec. V-C (R-Campus))
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| RESPLE LO本方法原文提出 | 0.28 m | (Cao et al., 2025, Sec. V-C) |
| RESPLE LIO本方法原文提出 | 0.27 m | (Cao et al., 2025, Sec. V-C) |
| CTE-MLO | 0.3 m | (Cao et al., 2025, Sec. V-C) |
| FAST-LIO2 | 2.7 m | (Cao et al., 2025, Sec. V-C) |
| Traj-LO | 80.31 m | (Cao et al., 2025, Sec. V-C) |
來源
Cao et al., 2025
(2025)RESPLE: Recursive Spline Estimation for LiDAR-Based OdometryIEEE Robotics and Automation Letters, 10(10):10666-10673
DOI 10.1109/lra.2025.3604758arXiv 2504.11580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:RESPLE (arXiv v3, marked as published in RA-L) https://arxiv.org/abs/2504.11580
- 程式碼釋出:ASIG-X/RESPLE https://github.com/ASIG-X/RESPLE
程式碼:https://github.com/ASIG-X/RESPLE(授權:GPL-3.0 (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。