In situ Fabricator BIM-referenced localization
作者為現地建造用的移動機械臂設計整合感測與控制系統。狀態估計以移動視窗估測器融合三件資訊：運動補償後的 VLP-16 掃描對建築模型網格取樣點雲的點到平面 ICP 位姿、IMU 以及輪式里程計，使機器人直接在建築模型座標中定位。接近施工位置時，系統再以末端執行器上三個互相正交的雷射測距儀，在多個臂姿下量測到附近牆面的距離，與模型網格光線追蹤出的平面比對，只最佳化基座位置以取得局部高精度定位。任務由 COMPAS 建築任務介面自動產生，並由全身 MPC 追蹤末端軌跡；最終點位以 Leica Nova TM50 全測站量測評估。
本頁內容
Mobile construction robot that localizes in the 3D building model by ICP of VLP-16 scans against a model-sampled cloud fused with IMU and wheel odometry in a moving horizon estimator, then refines its position near each task with end-effector laser distance sensors matched to model planes; end-effector dot positions checked with a Leica total station.
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 (Velodyne VLP-16)、IMU (Xsens MTi-100)、wheel encoders、three orthogonal laser distance sensors on the end-effector |
|---|---|
| 原文測試平台 | mobile manipulator (Inspector Bots Super Mega Bot base with Kinova Jaco 6-DoF arm, named Waco) |
| 狀態估計 | ConFusion moving horizon estimator fusing ICP pose updates against the building model, IMU and wheel odometry, with IMU forward propagation to compensate LiDAR latency; near task locations a separate high-accuracy localization (HAL) optimizes the static base position from laser distance measurements ray-traced against the 3D mesh, keeping orientation fixed (Sec. III-A, III-B) |
| 資料關聯 | point-to-plane ICP between motion-compensated LiDAR scans and a point cloud sampled from the 3D CAD triangle mesh; HAL finds the intersected mesh planes by ray tracing with a Cauchy robust cost (Sec. III-A, III-B) |
| 時間表示 | moving-horizon optimization over discrete states; IMU propagation between updates |
| 去畸變 | motion-compensated 3D LiDAR scans (method not detailed) (Sec. III-A) |
| 迴圈閉合 | none (localization in a known building model) |
| 全域最佳化 | none |
| 地圖表示 | point cloud sampled from the 3D CAD triangle mesh of the building model; Octomap occupancy map initialized from the model and updated from LiDAR scans for planning |
| 先驗資訊 | 3D building model (CAD triangle mesh from COMPAS or Rhino Grasshopper task interface); known approximate starting location |
| 可輸出幾何 | robot base and end-effector poses in the building-model frame; no point-cloud map product evaluated |
| 計算需求 | on-board Intel Core i7-6700 CPU @ 3.4 GHz with 16 GB RAM; whole-body MPC at about 100 Hz; off-board computer for the building task interface (Sec. III, IV-A) |
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 | Velodyne VLP-16 | 方法輸入 | 未標示 | 3D LiDAR for model-based ICP localization and obstacle updates | (Gawel et al., 2019, Sec. IV-A) |
| 慣性量測單元（IMU） | Xsens MTi-100 | 方法輸入 | 未標示 | process constraint in the MHE and latency compensation | (Gawel et al., 2019, Sec. IV-A, Sec. III-A3) |
| 輪式或腿式里程計 | wheel encoders | 方法輸入 | 未標示 | differential drive kinematic model; low confidence while wheels turn | (Gawel et al., 2019, Sec. III-A2, Sec. IV-A) |
| 全測站 | Leica Nova TM50 | 參考或真值量測 | 未標示 | measures commanded task locations and final end-effector placements | (Gawel et al., 2019, Sec. IV-B) |
| 載具平台 | Inspector Bots Super Mega Bot mobile base with Kinova Jaco arm (robot 'Waco') | 方法輸入 | 未標示 | skid-steered four-wheel base, 6-DoF arm, custom 3D-printed end-effector with spring-loaded marker | (Gawel et al., 2019, Sec. IV-A, Fig. 3) |
| 運算硬體 | Intel Core i7-6700 CPU @ 3.4 GHz, 16 GB RAM | 執行運算平台 | 未標示 | on-board computer; separate off-board computer over WLAN for the building task interface | (Gawel et al., 2019, Sec. IV-A) |
| 其他 | three orthogonal laser distance measurement sensors (sensor head at end-effector) | 方法輸入 | 未標示 | used for high-accuracy localization against model planes, six end-effector poses in the experiment | (Gawel et al., 2019, Sec. III-B, Sec. IV-A, Fig. 3) |
作者報告的優勢與限制
優勢
- Mean relative errors of 3.3 to 5.9 mm within the 9-dot patterns, supporting locally repeatable tasks (Table I; Sec. IV-C)
- Sub-cm absolute positioning (7.0 mm mean) at location A using only on-board sensing (Table I; Abstract)
- Reliable operation over 30 experiments including autonomous loops between task locations and re-planning around dynamic obstacles (Sec. IV-C)
限制
- Absolute error strongly depends on the task location: 32.9 mm and 26.1 mm mean at locations B and C (Table I)
- Distant lateral reference walls (11.5 m at B) magnify small rotational calibration errors of the laser distance sensors (Sec. IV-C2)
- Floor inclination at C and brick walls not represented in the building model add error; a simplified model increased relative errors at C by factors of 2 to 3 (Sec. IV-C2)
- HAL orientation estimate was less reliable than the LiDAR estimate and is kept fixed; the HAL routine is hand-designed (Sec. III-B, V)
- Global localization is out of scope; an approximate initial guess from the known start is needed (Sec. III-A)
營建工程相關證據
實驗在作者描述為接近真實施工條件的環境中進行（未完成的混凝土牆、少量雜物、模型與現況有偏差），並以 Leica Nova TM50 全測站量測末端執行器打點位置作為參考；文中未說明是否為施工中工地，所以歸為類工地的受控實驗（Sec. IV-B）。結果顯示以建築模型為參考的定位精度，會受到參考牆距離與模型未反映的現況偏差影響（Sec. IV-C）。
原文驗證環境：受控實驗、獨立參考量測、任務層驗證
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 7 筆紀錄。
Gawel et al., 2019 · Table I 本方法 6 筆
表格設定（擷取紀錄原文）：Autonomous approach from random locations several metres away, HAL with three laser distance sensors, then marking a 50 mm spaced 3 x 3 dot pattern on a wall; commanded and executed dot positions measured with a Leica Nova TM50 total station; distance from task location to the lateral HAL reference wall 1.7 m (A), 11.5 m (B), 4 m (C) (Gawel et al., 2019, Table I)
Mean absolute error (end-effector dot position, global coordinates w.r.t. building model origin)，authors' construction-like test site · task location A
這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Gawel et al., 2019 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| proposed system (LiDAR-to-model ICP + MHE + HAL + whole-body MPC)本方法原文提出 | 7 mm | (Gawel et al., 2019, Table I; Sec. IV-C) |
Gawel et al., 2019 · Text Sec.III 本方法 1 筆
指標MPC strategy running at approximately 100 Hz
資料集與序列不適用
表格設定（擷取紀錄原文）：Rate of the whole-body MPC that generates base and end-effector reference trajectories (Gawel et al., 2019, Text Sec.III)
MPC strategy running at approximately 100 Hz，不適用
這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Gawel et al., 2019 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| whole-body MPC (OCS2)本方法原文提出硬體：原文未報告 (the paper lists an on-board Intel Core i7-6700 CPU @ 3.4 GHz with 16 GB RAM but does not state which computer runs the MPC) | 100 Hz有附註註記（擷取紀錄）：other: approximate ('~100 Hz') | (Gawel et al., 2019, Sec. III) |
來源
Gawel et al., 2019
（2019）A Fully-Integrated Sensing and Control System for High-Accuracy Mobile Robotic Building Construction2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 2300-2307
DOI 10.1109/iros40897.2019.8967733arXiv 1912.01870
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：arXiv 1912.01870 v1 https://arxiv.org/abs/1912.01870
- repository copy：ETH Research Collection record https://doi.org/10.3929/ethz-b-000386713