Holistic Fusion
Holistic Fusion 是以 GTSAM 因子圖為核心的通用狀態估測框架：IMU 為骨幹，外部模組提供的位姿、位置、速度與地標量測都可作為因子接入；各個參考座標系（例如會漂移的 LiDAR 地圖座標系、里程計座標系與 GNSS 世界座標系）之間的對齊關係被當成隨機漫步的狀態一起最佳化，並沿路徑以局部關鍵影格對齊。線上以固定延遲平滑器提供 IMU 頻率的輸出，並另外提供平滑、不跳動的里程計座標系；離線則做整段批次最佳化，可作為事後處理的參考軌跡。
本頁內容
A task- and setup-agnostic GTSAM factor-graph framework that treats reference-frame alignments (drifting map, odometry and GNSS frames) as random-walk states, fuses heterogeneous absolute, relative, velocity and landmark measurements around IMU pre-integration, and provides online fixed-lag, smooth odometry and offline batch estimates.
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 | IMU (core)、LiDAR registration poses (e.g., Open3D SLAM, CompSLAM, Coin-LIO outputs)、GNSS (single or dual antenna)、leg kinematics or wheel encoder、mm-wave radar velocity、cabin rotary encoder (HEAP) |
|---|---|
| 原文測試平台 | ANYmal quadruped (hikes, parkour, indoor)、RACER Polaris RZR S4 1000 Turbo off-road vehicle、HEAP hydraulic walking excavator |
| 狀態估計 | GTSAM factor graph fusing IMU pre-integration with absolute pose, absolute position (GNSS), 3D landmark, local velocity and relative measurements; dynamic reference-frame alignment states modelled as an SE(3) random walk, local keyframe alignment along the path, landmark and global calibration states; online iSAM2 fixed-lag smoother with IMU-rate prediction and asynchronous updates, and offline batch optimization for post-processed ground truth (Sec. IV) |
| 資料關聯 | front-end agnostic: consumes poses, positions, velocities and landmarks from external modules (e.g., LiDAR scan-to-map registration, leg odometry, GNSS); outlier handling by robust norms (Sec. IV) |
| 時間表示 | states created at a configurable rate (10 to 100 Hz evaluated) with IMU-rate propagation; delayed and out-of-order measurements attached to the nearest states (Secs. IV-V; Table VII) |
| 去畸變 | not applicable (no raw point clouds are processed by the framework) |
| 迴圈閉合 | not part of the framework; handled by upstream modules |
| 全域最佳化 | offline batch optimization over the full mission; online fixed-lag smoothing (Sec. IV) |
| 地圖表示 | none internally; aligns external map frames (e.g., LiDAR map, odometry frame) to the world frame |
| 先驗資訊 | sensor extrinsics (optionally estimated as global states) |
| 可輸出幾何 | robot state in world, map and smooth odometry frames at IMU rate; offline trajectory used as post-processed ground truth |
| 計算需求 | evaluation PC with Intel i9 13900K; online latency about 22 to 30 us and asynchronous optimization 1.29 to 8.96 ms depending on state-creation rate (Table VII); offline optimization 0.66 s to 64.1 s per mission (Table IV) |
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 | VLP-16歸入：Velodyne VLP-16 | 方法輸入 | ANYmal parkour (authors) | LiDAR on ANYmal in the parkour experiment; its accumulated cloud meshed with HF offline poses | (Nubert et al., 2026, Sec. VI-C2; Fig. 18) |
| LiDAR | OS0-128歸入：Ouster OS0-128 | 方法輸入 | HEAP construction missions (authors) | LiDAR on HEAP; feeds CompSLAM scan-to-map registration and Coin-LIO | (Nubert et al., 2026, Sec. VI-E) |
| 載具平台 | ANYmal | 方法輸入 | ANYmal hike, parkour and indoor missions (authors) | quadruped with IMU, LiDAR, leg kinematics and a single GNSS antenna | (Nubert et al., 2026, Table III; Sec. VI-A) |
| 載具平台 | Polaris RZR S4 1000 Turbo (customized, RACER) | 方法輸入 | RACER off-road dataset | three LiDARs, mm-wave radar, single GNSS antenna and wheel encoder; 4.1 km at up to 9.66 m/s | (Nubert et al., 2026, Sec. VI-D) |
| 載具平台 | HEAP hydraulic walking excavator | 方法輸入 | HEAP construction missions (authors) | IMU, two GNSS antennas, cabin-to-chassis rotary encoder | (Nubert et al., 2026, Sec. VI-E) |
| 運算硬體 | PC with Intel i9 13900K | 執行運算平台 | 未標示 | all evaluations | (Nubert et al., 2026, Sec. VI-B) |
| 其他 | Qualisys mocap system | 參考或真值量測 | ANYmal parkour and indoor (authors) | ground truth for parkour and indoor experiments | (Nubert et al., 2026, Sec. VI-B) |
作者報告的優勢與限制
優勢
- On the two ANYmal hikes, HF World with GNSS filtering reaches ATE 0.49 and 0.12 m against post-processed ground truth, while TSIF odometry drifts to 33.38 and 18.22 m (Table V)
- HF Odom gives zero jumps (NOJ 0) and much lower jitter than world-frame estimates (Table VI)
- On five indoor mocap sequences, HF World with LiDAR registration and tight leg kinematics gives ATE 2.8 cm, comparable to Open3D-SLAM (3.0 cm) but at high rate (Table IX)
- Online latency in the lower microsecond range and asynchronous optimization of 1.29 to 8.96 ms for 10 to 100 Hz state creation (Table VII)
- On a walking excavator at an industrial construction site in Oberglatt, the mission could be completed although the corridor between two buildings was both GNSS-denied and geometrically degenerate for CompSLAM scan-to-map registration (Sec. VI-E1; Figs. 23-24)
限制
- Initialization must be handled carefully (risk of ill-posed problems or wrong local minima); with a single GNSS antenna yaw is initially unobservable; many tuning parameters for new users (Sec. VII-B)
- Hike, RACER and HEAP references are post-processed trajectories produced by offline batch optimization with GNSS, i.e., by the framework itself, not an independent system (Sec. VI-B)
- HEAP construction-site results are qualitative (Figs. 23-25)
- Leg-only HF estimates are less accurate than the engineered TSIF estimator (Tables VIII-IX)
營建工程相關證據
本文直接在施工機具上驗證：HEAP 液壓步行挖掘機於瑞士 Oberglatt 工業施工現場作業，建物之間的走廊同時失去 GNSS 且 LiDAR 幾何退化，框架透過同時融合 CompSLAM 掃描對地圖與 Coin-LIO 並估計各地圖座標系的漂移，仍完成任務並建立地圖。但施工現場部分只有定性結果（地圖與漂移曲線），定量評估來自四足機器人資料；其參考座標系對齊思路對長時間、跨區段的工地點雲整合具參考價值（推論）。
原文驗證環境：施工中工地、獨立參考量測、跨場域、受控實驗
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 99 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。
Nubert et al., 2026 · Table V 本方法 32 筆
表格設定（擷取紀錄原文）：ANYmal autonomous hikes (Forest and Mountain/Seealpsee); global ATE [m] and ARE [deg] against post-processed ground truth (offline HF batch optimization with GNSS); HF World = world frame, HF Odom = smooth odometry frame (Sec. IV-D3); LR = LiDAR registration; MINS diverged mid-way on Mountain ('div.'), 'split' excludes the divergence (Nubert et al., 2026, Table V)
ATE [m]，ANYmal hike missions (authors) · Forest
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 Nubert 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(Nubert et al., 2026, Table V)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| TSIF - Odom | 33.38 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| Open3D SLAM - LR | 1.46 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| MINS - World (div.) | 0.44 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| MINS - World (split) | 0.44 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| HF GNSS+IMU - World本方法原文提出 | 0.53 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| HF GNSS+IMU - Odom本方法原文提出 | 101.55 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| HF (LR-between) - World本方法原文提出 | 0.52 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| HF (LR-between) - Odom本方法原文提出 | 66 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| HF - World本方法原文提出 | 0.42 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| HF - Odom本方法原文提出 | 22.91 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| HF (GNSS filtered) - World本方法原文提出 | 0.49 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| HF (GNSS filtered) - Odom本方法原文提出 | 15.94 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
Nubert et al., 2026 · Table VI 本方法 24 筆
資料集與序列ANYmal hike missions (authors) · Forest and Mountain (average)
表格設定（擷取紀錄原文）：Average local estimation quality and smoothness over the two hikes; RTE and RRE averaged over all 1 m pairs; NOJ = jumps above 10 cm between consecutive 400 Hz estimates; jitter = jerk (third derivative of position); MINS full Forest plus split Mountain (Nubert et al., 2026, Table VI)
RTE [%]，ANYmal hike missions (authors) · Forest and Mountain (average)
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 Nubert 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(Nubert et al., 2026, Table VI)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| TSIF - Odom | 5.01% | (Nubert et al., 2026, Table VI; Sec. VI-C1) |
| Open3D-SLAM - LR (<=10 Hz) | 6.28% | (Nubert et al., 2026, Table VI; Sec. VI-C1) |
| MINS - World | 4.52% | (Nubert et al., 2026, Table VI; Sec. VI-C1) |
| HF (LR-between) - World本方法原文提出 | 4.84% | (Nubert et al., 2026, Table VI; Sec. VI-C1) |
| HF (LR-between) - Odom本方法原文提出 | 4.09% | (Nubert et al., 2026, Table VI; Sec. VI-C1) |
| HF - World本方法原文提出 | 4.07% | (Nubert et al., 2026, Table VI; Sec. VI-C1) |
| HF - Odom本方法原文提出 | 3.13% | (Nubert et al., 2026, Table VI; Sec. VI-C1) |
| HF (GNSS filtered) - World本方法原文提出 | 2.62% | (Nubert et al., 2026, Table VI; Sec. VI-C1) |
| HF (GNSS filtered) - Odom本方法原文提出 | 3.1% | (Nubert et al., 2026, Table VI; Sec. VI-C1) |
Nubert et al., 2026 · Table IX 本方法 20 筆
資料集與序列ANYmal indoor locomotion dataset (authors) · five sequences (average)
表格設定（擷取紀錄原文）：ANYmal indoor locomotion dataset, five sequences, averaged; Qualisys mocap ground truth; mean values stored, standard deviation in the outcome field; LR = LiDAR registration (Open3D-SLAM) (Nubert et al., 2026, Table IX)
ATE [cm]，ANYmal indoor locomotion dataset (authors) · five sequences (average)
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 Nubert 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(Nubert et al., 2026, Table IX)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Open3D-SLAM (LR only, low rate) | 3 cm有附註註記（擷取紀錄）：std as printed: 1.5 | (Nubert et al., 2026, Table IX; Sec. VI-C3) |
| HF World (IMU + leg odometry velocity, loosely fused)本方法原文提出 | 13.7 cm有附註註記（擷取紀錄）：std as printed: 6.3 | (Nubert et al., 2026, Table IX; Sec. VI-C3) |
| TSIF (IMU + leg kinematics, tightly fused) | 7.9 cm有附註註記（擷取紀錄）：std as printed: 2.9 | (Nubert et al., 2026, Table IX; Sec. VI-C3) |
| HF World (IMU + leg kinematics, tightly fused)本方法原文提出 | 10.1 cm有附註註記（擷取紀錄）：std as printed: 4.0 | (Nubert et al., 2026, Table IX; Sec. VI-C3) |
| HF World (IMU + LR + leg odometry velocity, loosely fused)本方法原文提出 | 4 cm有附註註記（擷取紀錄）：std as printed: 2.4 | (Nubert et al., 2026, Table IX; Sec. VI-C3) |
| HF World (IMU + LR + leg kinematics, tightly fused)本方法原文提出 | 2.8 cm有附註註記（擷取紀錄）：std as printed: 1.4 | (Nubert et al., 2026, Table IX; Sec. VI-C3) |
| HF Offline (IMU + LR + leg kinematics, tightly fused)本方法原文提出 | 2.8 cm有附註註記（擷取紀錄）：std as printed: 1.4 | (Nubert et al., 2026, Table IX; Sec. VI-C3) |
Nubert et al., 2026 · Table VII 本方法 12 筆
資料集與序列ANYmal hike missions (authors) · hike
表格設定（擷取紀錄原文）：ANYmal hike: computational complexity and accuracy versus state-creation rate; mean values stored, standard deviation in the outcome field; evaluation PC Intel i9 13900K (Nubert et al., 2026, Table VII)
Latency [us]，ANYmal hike missions (authors) · hike
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 Nubert 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(Nubert et al., 2026, Table VII)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| HF at 10 Hz state creation本方法原文提出硬體：PC with Intel i9 13900K | 22.41 us有附註註記（擷取紀錄）：std as printed: 13.62 | (Nubert et al., 2026, Table VII; Sec. VI-C1) |
| HF at 40 Hz state creation本方法原文提出硬體：PC with Intel i9 13900K | 25.31 us有附註註記（擷取紀錄）：std as printed: 17.07 | (Nubert et al., 2026, Table VII; Sec. VI-C1) |
| HF at 100 Hz state creation本方法原文提出硬體：PC with Intel i9 13900K | 29.63 us有附註註記（擷取紀錄）：std as printed: 19.99 | (Nubert et al., 2026, Table VII; Sec. VI-C1) |
其他比較組
列出其餘 1 個比較組
來源
Nubert et al., 2026
（2026）Holistic Fusion: Task- and Setup-Agnostic Robot Localization and State Estimation With Factor GraphsIEEE Transactions on Robotics, 42, pp. 3366-3387
DOI 10.1109/tro.2026.3714645arXiv 2504.06479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：Holistic Fusion arXiv v2 (accepted version) https://arxiv.org/abs/2504.06479v2
- 程式碼釋出：leggedrobotics/holistic_fusion https://github.com/leggedrobotics/holistic_fusion
- successor：Generalizes the dual-graph estimator Graph MSF built for construction robots (see Nubert et al., 2022a) not_applicable
程式碼：https://github.com/leggedrobotics/holistic_fusion（授權：BSD-3-Clause (GitHub license metadata)）。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。