RKO-LIO
RKO-LIO 不採用卡爾曼濾波或預積分因子圖,而是假設相鄰 LiDAR 幀間線加速度與角速度固定,以簡化模型積分 IMU 取得 ICP 初值與逐點去畸變,再以掃描對地圖 ICP 精修。作者在 ICP 中加入依 IMU 加速度資訊自適應調整的姿態正則化,並主張不需感測器特定的雜訊模型與校正,即可用同一組參數跨車載、背包、四足與無人機平台。
本頁內容
LiDAR-inertial odometry without filtering or preintegration: a simple constant-acceleration IMU model provides the ICP initial guess and deskew, and an adaptive IMU-based orientation regularizer is added to scan-to-map ICP, run with one configuration across platforms.
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、IMU (consumer to industrial grade) |
|---|---|
| 原文測試平台 | vehicle (own car, HeLiPR)、backpack (Oxford Spires, DigiForests)、legged (Unitree Go1)、UAV (DJI M210 v2, DRZ Living Lab)、tree-harvesting machine (qualitative, Fig. 1) |
| 狀態估計 | No filter or factor graph for the pose: IMU samples between two scans are bias- and gravity-compensated, averaged, and integrated with a constant linear acceleration and angular velocity model to give the ICP initial guess and per-point deskew; scan-to-map point-to-point ICP adds an accelerometer-based orientation cost weighted by 1/beta, with beta = beta0 (1 + sigma_a^2) and beta0 = 200, where body acceleration comes from a small Kalman filter with maximum expected jerk 3 m/s3; biases assumed constant and estimated from the first inter-scan interval |
| 資料關聯 | Scan-to-map ICP building on the KISS-ICP scan-alignment module: point-to-point residuals with a fixed association threshold of 0.5 m in a VDB voxel map (1.0 m voxels), double downsampling at 0.5 v for map update and 1.5 v for registration (optional off switch for sparse LiDARs), points within 1 m of the sensor clipped |
| 時間表示 | discrete poses with per-point deskew from IMU-derived relative transforms |
| 去畸變 | per-point transform from IMU-based motion between LiDAR frames (Sec. III-B) |
| 迴圈閉合 | none |
| 全域最佳化 | none |
| 地圖表示 | VDB voxel grid (voxel size 1.0 m) storing a fixed number of points per voxel; double downsampling |
| 先驗資訊 | none |
| 可輸出幾何 | odometry and local voxel map; export format 原文未報告 |
| 計算需求 | 原文未報告: no hardware or per-scan timings are given; the authors only state that the system runs faster than the sensor frame rate on all presented datasets |
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 | Hesai QT64 | 資料集感測器 | Oxford Spires | backpack-mounted | (Malladi et al., 2026, Sec. IV-A; Fig. 1) |
| LiDAR | Velodyne VLP-16 | 資料集感測器 | Leg-KILO dataset | on a Unitree Go1 quadruped | (Malladi et al., 2026, Sec. IV-A; Fig. 1) |
| LiDAR | Aeva Aeries II | 資料集感測器 | HeLiPR | low field-of-view solid-state LiDAR, one of four LiDARs in the dataset | (Malladi et al., 2026, Sec. IV-A; Sec. IV-B) |
| LiDAR | Livox Avia | 資料集感測器 | HeLiPR | non-repetitive scan pattern | (Malladi et al., 2026, Sec. IV-A) |
| LiDAR | Hesai XT32, QT32 and QT64 (different sessions) | 資料集感測器 | DigiForests | backpack sensor rig; LiDAR inclined 45 deg in the first season | (Malladi et al., 2026, Sec. IV-A; Sec. IV-B) |
| LiDAR | Ouster OS-0歸入:Ouster OS0 | 資料集感測器 | DRZ Living Lab | mounted on a drone | (Malladi et al., 2026, Sec. IV-A) |
| LiDAR | OS1-128歸入:Ouster OS1-128 | 方法輸入 | own car dataset | car-mounted | (Malladi et al., 2026, Sec. IV-A; Fig. 1) |
| LiDAR | Hesai XT32歸入:Hesai XT-32 | 方法輸入 | 未標示 | on a tree-harvesting machine (qualitative example only) | (Malladi et al., 2026, Fig. 1) |
| 地面雷射掃描儀(TLS) | terrestrial laser-scanning map (scanner model not reported) | 參考或真值量測 | Oxford Spires | each undistorted scan registered to the TLS map to compute ground truth | (Malladi et al., 2026, Sec. IV-A) |
| 慣性量測單元(IMU) | cellphone-grade IMU (Fig. 1 caption names an Alphasense IMU) | 資料集感測器 | Oxford Spires | 400 Hz | (Malladi et al., 2026, Sec. IV-A; Fig. 1) |
| 慣性量測單元(IMU) | onboard IMU of Unitree Go1 | 資料集感測器 | Leg-KILO dataset | 500 Hz | (Malladi et al., 2026, Sec. IV-A) |
| 慣性量測單元(IMU) | Xsens MTi-300 | 資料集感測器 | HeLiPR | 100 Hz | (Malladi et al., 2026, Sec. IV-A) |
| 慣性量測單元(IMU) | built-in InvenSense IMU of the LiDAR | 方法輸入 | own car dataset | 100 Hz | (Malladi et al., 2026, Sec. IV-A; Fig. 1) |
| 慣性量測單元(IMU) | Xsens MTi-100 | 方法輸入 | 未標示 | on a tree-harvesting machine (qualitative example only) | (Malladi et al., 2026, Fig. 1) |
| GNSS 接收器 | RTK-GPS INS based system | 參考或真值量測 | HeLiPR | ground truth trajectories for each sensor | (Malladi et al., 2026, Sec. IV-A) |
| GNSS 接收器 | SBG Ellipse-D GNSS-INS | 參考或真值量測 | own car dataset | reference poses from offline LiDAR bundle adjustment incorporating RTK-GPS | (Malladi et al., 2026, Sec. IV-A) |
| 載具平台 | Unitree Go1 quadruped | 資料集感測器 | Leg-KILO dataset | indoor sequences, one parking lot and one running sequence | (Malladi et al., 2026, Sec. IV-A) |
| 載具平台 | DJI M210 v2 | 資料集感測器 | DRZ Living Lab | drone platform | (Malladi et al., 2026, Sec. IV-A) |
| 其他 | motion capture system | 參考或真值量測 | DRZ Living Lab | ground truth for nine sequences | (Malladi et al., 2026, Sec. IV-A) |
作者報告的優勢與限制
優勢
- Same configuration across Oxford Spires, Leg-KILO, HeLiPR, DigiForests, DRZ Living Lab and own car datasets (Sec. IV)
- Generally best odometry on both metrics on Oxford Spires and mostly on par with a reference SLAM result (Sec. IV-B)
- Only compared method without failures on all presented datasets (Sec. IV-B)
- Rural 52 km sequence ATE 714.82 m versus 1086.51 m for FAST-LIO2 (Table II)
限制
- Relies on the constant linear acceleration and angular velocity assumption between frames; authors bound its error for typical 0.1 s intervals (Sec. III-B)
- Odometry only; no loop closure (inference from system scope)
- Biases are assumed constant and initialized assuming no motion over the first interval; initialization was disabled on Oxford Spires because sequences start in motion (Sec. III-B; Sec. IV-A)
- Not best everywhere: FAST-LIO2 better on Residential and HeLiPR Bridge, Leg-KILO better on quadruped data (Tables II, III, V)
- Disabling double downsampling for sparse LiDARs improved results, which is a sensor-aware setting (Sec. IV-C)
營建工程相關證據
Oxford Spires 以背包式 LiDAR 在大學校園室內外錄製,地面真值由逐掃描對準地面雷射掃描(TLS)地圖產生,屬既有建築情境;未於施工現場測試。
原文驗證環境:公開基準、獨立參考量測、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 5 個比較組,合計 39 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 1 組列在最後,並連到性能比較頁。
Malladi et al., 2026 · Table II 本方法 15 筆
指標ATE (m)
表格設定(擷取紀錄原文):Own car sequences (OS1-128 with built-in InvenSense IMU; Forest 20 km, Rural 52 km); reference by offline LiDAR bundle adjustment with RTK-GPS; ATE only transcribed; ablation rows no-AVG no-AR and no-AR disable IMU averaging and adaptive regularization; dash means failure to run (Malladi et al., 2026, Table II)
ATE (m),own car dataset · Urban
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 Malladi et al., 2026 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(Malladi et al., 2026, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| KISS-ICP | 8.17 m | (Malladi et al., 2026, Table II) |
| DLIO | 4.94 m | (Malladi et al., 2026, Table II) |
| FAST-LIO2 | 3.53 m | (Malladi et al., 2026, Table II) |
| Ours (RKO-LIO)本方法原文提出 | 3.52 m | (Malladi et al., 2026, Table II) |
| Ours, no-AVG, no-AR (ablation)本方法 | 4.16 m | (Malladi et al., 2026, Table II) |
| Ours, no-AR (ablation)本方法 | 5 m | (Malladi et al., 2026, Table II) |
Malladi et al., 2026 · Table I 本方法 10 筆
表格設定(擷取紀錄原文):Oxford Spires backpack (Hesai QT64); ground truth by registering undistorted scans to a TLS map; averages over all sequences of each scene; odometry without loop closure except the VILENS-SLAM reference; initialization disabled because sequences start in motion (Malladi et al., 2026, Table I)
ATE (m), averaged over sequences of each scene,Oxford Spires · Blenheim
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 Malladi et al., 2026 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(Malladi et al., 2026, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| KISS-ICP | 1.16 m | (Malladi et al., 2026, Table I) |
| DLIO | 11.35 m | (Malladi et al., 2026, Table I) |
| FAST-LIO2 | 0.94 m | (Malladi et al., 2026, Table I) |
| Ours (RKO-LIO)本方法原文提出 | 0.2 m | (Malladi et al., 2026, Table I) |
| VILENS-SLAM (SLAM reference, results from Tao et al.) | 0.56 m | (Malladi et al., 2026, Table I) |
Malladi et al., 2026 · Table III 本方法 6 筆
表格設定(擷取紀錄原文):Leg-KILO dataset, Unitree Go1 quadruped with VLP-16 and 500 Hz IMU; Indoor ground truth from a prior map, others from offline optimization with loop closures; Leg-KILO also uses leg kinematics (Malladi et al., 2026, Table III)
ATE (m),Leg-KILO dataset · Corridor
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 Malladi et al., 2026 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(Malladi et al., 2026, Table III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| KISS-ICP | 10.59 m | (Malladi et al., 2026, Table III) |
| DLIO | 2.7 m | (Malladi et al., 2026, Table III) |
| FAST-LIO2 | 0.28 m | (Malladi et al., 2026, Table III) |
| Leg-KILO | 0.17 m | (Malladi et al., 2026, Table III) |
| Ours (RKO-LIO)本方法原文提出 | 0.22 m | (Malladi et al., 2026, Table III) |
Malladi et al., 2026 · Table IV 本方法 6 筆
表格設定(擷取紀錄原文):DigiForests backpack sessions (Hesai XT32, QT32, QT64; LiDAR inclined 45 deg in the first season); reference trajectories from offline VILENS with GNSS and loop closures; averages per season; dash means failure on at least one sequence of that season (Malladi et al., 2026, Table IV)
ATE (m), season average,DigiForests · 2023-03
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。
- 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 Malladi et al., 2026 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(Malladi et al., 2026, Table IV)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| KISS-ICP | 無數值失敗註記(擷取紀錄):failed on at least one sequence of the recording period (dash) | (Malladi et al., 2026, Table IV) |
| DLIO | 0.27 m | (Malladi et al., 2026, Table IV) |
| FAST-LIO2 | 無數值失敗註記(擷取紀錄):failed on at least one sequence of the recording period (dash) | (Malladi et al., 2026, Table IV) |
| Ours (RKO-LIO)本方法原文提出 | 0.18 m | (Malladi et al., 2026, Table IV) |
其他比較組
來源
Malladi et al., 2026
(2026)Robust Approach for LiDAR-Inertial Odometry Without Sensor-Specific ModelingIEEE Robotics and Automation Letters, 11(6):7420-7427
DOI 10.1109/lra.2026.3685966arXiv 2509.06593程式碼
同儕審查已出版已讀全文近十年
相關版本
- 預印本:A Robust Approach for LiDAR-Inertial Odometry Without Sensor-Specific Modeling (arXiv v2) https://arxiv.org/abs/2509.06593
- 程式碼釋出:PRBonn/rko_lio https://github.com/PRBonn/rko_lio
程式碼:https://github.com/PRBonn/rko_lio(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。