Pronto
Pronto 是為腿式機器人設計的模組化擴展卡爾曼濾波器：以 IMU 作為高頻過程模型，先融合腿部運動學與接觸偵測得到的速度，再把延遲且低頻的視覺里程計與 LiDAR 點雲配準結果，以鬆耦合的位姿修正方式插入約 10 秒的量測歷史中重新傳播。這樣可在控制迴路中提供 250 至 1000 Hz 的低延遲狀態估測，同時利用外感測器抑制長時間漂移。
本頁內容
Modular EKF for legged robots that runs an IMU process model with leg-odometry velocity updates at control rate and inserts delayed, low-rate visual odometry and ICP-based LiDAR corrections into a stored measurement history, giving low-latency, low-drift estimates for closed-loop locomotion.
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 (KVH 1750, KVH 1775, 3DM-GX4-25 or Xsens MTi-100 depending on robot)、joint encoders and force or torque sensing for leg odometry、stereo camera (Carnegie Robotics Multisense SL) or RGB-D camera (Intel RealSense D435) for FOVIS visual odometry、LiDAR (Hokuyo UTM-30LX-EW spinning in the Multisense SL, or Velodyne VLP-16 on ANYmal) for AICP registration |
|---|---|
| 原文測試平台 | Atlas humanoid、Valkyrie humanoid、HyQ quadruped、ANYmal quadruped |
| 狀態估計 | modular extended Kalman filter with an IMU process model on the real-time control computer; leg-odometry velocity updates; zero-velocity gyro bias update when stationary at least 400 ms; loosely coupled pose or position corrections from FOVIS visual odometry and from AICP LiDAR registration applied through a measurement history of typically 10 s to handle their latency (Secs. 4-5) |
| 資料關聯 | leg odometry from kinematics with contact detection (Schmitt trigger on humanoids, probabilistic ground-reaction-force contact on quadrupeds); FOVIS feature-based stereo or RGB-D visual odometry; AICP auto-tuned ICP with planar pre-filtering and overlap-based outlier rejection against a reference cloud (Sec. 4) |
| 時間表示 | discrete EKF at the IMU rate (250 to 1000 Hz) with delayed asynchronous exteroceptive updates at 1 to 15 Hz inserted into the stored history and re-propagated (Sec. 5.1; Fig. 6) |
| 去畸變 | not described in the paper; AICP registers accumulated point clouds |
| 迴圈閉合 | no |
| 全域最佳化 | none; the filter marginalizes previous states (Sec. 9) |
| 地圖表示 | AICP reference point cloud updated after a travelled distance; no global map is maintained by the estimator |
| 先驗資訊 | robot kinematic model; fixed measurement covariances per module; for HyQ outdoor ground truth an ICP prior map was used for evaluation only |
| 可輸出幾何 | base pose and velocity at control rate for closed-loop locomotion |
| 計算需求 | consumer-grade processors (e.g., laptop-class Intel i7), no GPU; proprioceptive thread on the control computer, VO and LiDAR on a perception computer (Sec. 5) |
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 | UTM-30LX-EW (spinning, in Multisense SL)歸入：Hokuyo UTM-30LX | 方法輸入 | Atlas, Valkyrie, HyQ | 40 Hz; full-rotation FoV 220 x 180 deg (Atlas entry); produces sparse clouds that AICP accumulates | (Camurri et al., 2020, Table 1) |
| LiDAR | VLP-16歸入：Velodyne VLP-16 | 方法輸入 | ANYmal | mounted on top of ANYmal for localization and global mapping; used by AICP | (Camurri et al., 2020, Sec. 6.4) |
| 慣性量測單元（IMU） | KVH 1750 | 方法輸入 | Atlas (DRC Finals) | 333 Hz; initial bias 0.5 deg/h and 0.5 mg; bias stability 0.05 deg/h and 0.05 mg | (Camurri et al., 2020, Table 1) |
| 慣性量測單元（IMU） | 3DM-GX4-25 | 方法輸入 | Valkyrie | 500 Hz | (Camurri et al., 2020, Table 1) |
| 慣性量測單元（IMU） | KVH 1775 | 方法輸入 | HyQ | 1000 Hz; initial bias 0.5 deg/h and 0.5 mg; bias stability 0.05 deg/h and 0.05 mg | (Camurri et al., 2020, Table 1) |
| 慣性量測單元（IMU） | MTi-100歸入：Xsens MTi-100 | 方法輸入 | ANYmal | 400 Hz; initial bias 0.2 deg/s and 5 mg; bias stability 10 deg/h and 15 mg | (Camurri et al., 2020, Table 1) |
| 雙目相機 | Multisense SL | 方法輸入 | Atlas, Valkyrie, HyQ | tri-modal ruggedized sensor; stereo 10 Hz, 1024 x 1024 px, 80 x 80 deg FoV (CMV4000 imager) on Atlas; used on Atlas, Valkyrie and HyQ | (Camurri et al., 2020, Table 1; Sec. 6) |
| RGB-D 相機 | RealSense D435歸入：Intel RealSense D435 | 方法輸入 | ANYmal | 30 Hz, 848 x 480 for VO; a second downward-facing D435 also mounted | (Camurri et al., 2020, Table 1; Sec. 7.5) |
| 輪式或腿式里程計 | AEDA3300-BE1 | 方法輸入 | HyQ | joint encoders, 1,000 Hz, resolution <0.0045 deg; leg odometry input | (Camurri et al., 2020, Table 1) |
| 輪式或腿式里程計 | ANYdrive | 方法輸入 | ANYmal | joint encoders (resolution <0.025 deg) and torque at 400 Hz; leg odometry input | (Camurri et al., 2020, Table 1) |
| 全測站 | TS16 | 參考或真值量測 | ANYmal (Fire Service College) | laser tracking system following a reflective prism with millimetre accuracy; ANYmal experiment | (Camurri et al., 2020, Sec. 7.1.1) |
| 其他 | Omega85 | 方法輸入 | Valkyrie | foot force and torque sensing for contact detection | (Camurri et al., 2020, Table 1) |
| 其他 | Vicon | 參考或真值量測 | Valkyrie, HyQ | 100 Hz, millimetre-accurate; Valkyrie and HyQ indoor experiments | (Camurri et al., 2020, Sec. 7.1.1) |
| 其他 | Burster 8417 | 方法輸入 | HyQ | force sensing, 1,000 Hz, resolution <25 N | (Camurri et al., 2020, Table 1) |
論文圖片
只收錄原文以開放授權（open license）釋出的圖片，並依授權條件標示出處、圖號、授權與修改方式。
Figure 5系統方塊圖：IMU 過程模型與腿部里程計在控制電腦即時執行，視覺里程計與 LiDAR 配準在感知電腦執行後回饋修正
出處：Camurri et al., 2020，Figure 5。授權：CC BY 4.0。原始圖檔。修改：轉存為 WebP 格式。
Figure 7Atlas 在 DRC 決賽資料上的 AICP 表現：上方為 206 片點雲對齊結果的上視圖，左為含人群的原始點雲，右為濾除後點雲；下方比較未套用修正時閥門在各片點雲中位置不一致，以及定位成功後估測一致的結果
出處：Camurri et al., 2020，Figure 7。授權：CC BY 4.0。原始圖檔。修改：轉存為 WebP 格式。
Figure 9HyQ 室內重複性測試、室外昏暗工業區探索測試，以及不同感測組合估測軌跡的比較
出處：Camurri et al., 2020，Figure 9。授權：CC BY 4.0。原始圖檔。修改：轉存為 WebP 格式。
Figure 12ANYmal 在消防訓練場的實驗：場景照片、機上相機畫面（水面反光），以及估測軌跡與真值比較
出處：Camurri et al., 2020，Figure 12。授權：CC BY 4.0。原始圖檔。修改：轉存為 WebP 格式。
作者報告的優勢與限制
優勢
- Validated on four legged robots over more than 2 h and 1.37 km, including closed-loop control during dynamic gaits (Sec. 7; Table 2)
- RPE over 10 m of 0.016 to 0.033 m in the smaller humanoid and HyQ experiments (Table 2)
- On Valkyrie the translation error averaged 1.6 cm and never exceeded 7.5 cm, while the estimator without LiDAR drifted without bound (Sec. 7.3)
- On ANYmal in an industrial fire-training site, fusing VO and AICP reduced RPE by 60 % compared with IMU and leg odometry only; the pose was within 30 cm of ground truth after 240 m (Sec. 7.5)
- Runs on laptop-class CPUs without GPU and is released as open-source ROS packages (Sec. 5)
限制
- No triaging when exteroceptive sources disagree; measurement confidence is a fixed covariance (Sec. 8)
- Loosely coupled design; the authors name tightly coupled joint optimization as future work (Sec. 8)
- No loop closure or global map; ground truth was not available for all experiments (Table 2)
- The Atlas and HyQ outdoor accuracy values are bounds from indirect evaluation (Table 2 footnotes)
營建工程相關證據
腿式機器人（如 ANYmal、Spot）已用於工地巡檢與掃描，本文提供其底層狀態估測的完整設計與長時間實測。ANYmal 實驗在消防訓練場的工業設施內進行，有水面反光、低光照與崎嶇地形，並以 Leica TS16 全測站作為真值，與施工現場的條件部分相近（推論）。但本文只評估位姿，不產生或評估點雲地圖。
原文驗證環境：受控實驗、獨立參考量測、跨場域
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 2 個比較組，合計 14 筆紀錄。
Camurri et al., 2020 · Table 2 本方法 8 筆
表格設定（擷取紀錄原文）：Summary of experiments; RPE = translational part of relative pose error evaluated over 10 m distance; OL = online, CL = used in the control loop; GT = ground truth available (Vicon for Valkyrie and HyQ indoor, Leica TS16 for ANYmal) (Camurri et al., 2020, Table 2)
RPE [m], over 10 m，Pronto experiments (authors) · Exp. 7 (ANYmal)
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 Camurri et al., 2020 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(Camurri et al., 2020, Table 2)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Pronto (IMU+LO+VO+AICP, offline)本方法原文提出 | 0.34 m | (Camurri et al., 2020, Table 2) |
| Pronto (IMU+LO only (without VO and AICP), offline)本方法原文提出 | 0.83 m | (Camurri et al., 2020, Table 2) |
Camurri et al., 2020 · Sec. 7 text 本方法 6 筆
資料集與序列Pronto experiments (authors) · Exp. 6c-6e (HyQ)
表格設定（擷取紀錄原文）：Values stated in the text of Section 7 (Camurri et al., 2020, Sec. 7 text)
average 3D median translation error，Pronto experiments (authors) · Exp. 6c-6e (HyQ)
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 Camurri et al., 2020 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(Camurri et al., 2020, Sec. 7 text)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Pronto (IMU+LO+AICP)本方法原文提出 | 0.049 m有附註註記（擷取紀錄）：approximate as printed ('~4.9 cm') | (Camurri et al., 2020, Sec. 7.4; Fig. 11) |
| Pronto (IMU+LO+AICP+VO)本方法原文提出 | 0.032 m | (Camurri et al., 2020, Sec. 7.4; Fig. 11) |
來源
Camurri et al., 2020
（2020）Pronto: A Multi-Sensor State Estimator for Legged Robots in Real-World ScenariosFrontiers in Robotics and AI, 7, article 68
DOI 10.3389/frobt.2020.00068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
- 程式碼釋出：ori-drs/pronto (core EKF and leg odometry), with fovis_ros and aicp_mapping modules https://github.com/ori-drs/pronto
程式碼：https://github.com/ori-drs/pronto（授權：LGPL-2.1 (GitHub license metadata for ori-drs/pronto)）。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。