TL;DR
This paper introduces an adaptive robotic system for finishing concrete shield segments, overcoming the limitations of traditional methods by using 3D vision and point cloud data to dynamically adjust tool paths. This approach significantly improves finishing quality and efficiency, paving the way for more automated and precise construction processes.
What Problem Does This Paper Address?
The finishing of concrete shield segments, a crucial step in tunnel engineering, has historically relied on manual methods that are inefficient, labor-intensive, and struggle to achieve consistent quality, especially on complex curved surfaces. Existing robotic solutions often use pre-programmed, fixed trajectories, which fail to account for real-world variations like manufacturing tolerances, installation errors, or inherent material deformations. This lack of adaptability leads to inconsistent tool-to-surface contact and compromised finishing quality. The paper aims to bridge this gap by developing an intelligent robotic system that can perceive and adapt to the actual surface geometry in real-time.
Methods
The core of the proposed solution is an adaptive trajectory planning method integrated into a robotic system. This system comprises:
- Automated Guided Vehicle (AGV): For precise positioning of the robot at multiple workstations.
- 6-DOF Serial Manipulator: Providing the necessary reach and flexibility for finishing operations.
- 3D Vision System: Mounted on the robot's end-effector to capture detailed point cloud data of the segment surface.
The methodology involves several key steps:
- Offline Trajectory Programming: A "rectangular" trajectory, consisting of concentric paths, is initially planned in a simulated environment. This ensures comprehensive coverage of the segment surface.
- Point Cloud Acquisition and Processing: After the AGV positions the robot, the 3D camera captures multi-view point cloud data. This data is then registered, stitched, and fused to create a complete 3D model of the segment's surface.
- Surface Irregularity Analysis: A neural network (PointNet++ architecture) analyzes the fused point cloud to identify and quantify surface imperfections (protrusions/depressions) and their dimensions.
- Adaptive Trajectory Compensation: Based on the identified surface irregularities, a "look-ahead tilt angle" is calculated. This angle informs real-time adjustments to the robot's trajectory and the trowel's posture to maintain optimal contact with the local surface normal.
- Multi-Coordinate System Transformation: A robust framework is established to transform trajectory data accurately between different coordinate systems (mold, QR code, robot base) to ensure precise execution.
Datasets
Not explicitly discussed in the paper. The neural network for defect detection was trained on a dataset of 342 point cloud samples, augmented with transformations.
Key Insights and Ideas
- Bridging the Offline-Online Gap: The system effectively combines the benefits of offline programming (efficiency, pre-planning) with online adaptive capabilities (real-time perception and adjustment), a critical need for robust industrial automation.
- Perception-Driven Adaptation: The use of 3D vision and point cloud analysis to understand the "as-built" surface geometry is central to overcoming the limitations of fixed trajectories.
- Look-Ahead Tilt Angle: This novel concept allows the system to proactively compensate for surface undulations by adjusting the trowel's tilt angle before encountering the irregularity, rather than reactively.
- Robust Coordinate System Framework: The meticulous handling of multiple coordinate systems is essential for translating planned trajectories into precise robot movements in a dynamic industrial setting.
Key Results
The experimental validation demonstrated significant improvements:
- Surface Quality: The adaptive method reduced surface height variations along a representative arc from 38.2 mm (fixed-trajectory) to 4.7 mm, an improvement of 87.7%.
- Finishing Uniformity: Visual inspection and quantitative data confirmed a consistently flatter and more uniform surface finish compared to the conventional fixed-trajectory method.
- Accuracy: The neural network achieved a mean absolute error of 1.8 mm in defect center localization and 0.9 mm in dimensional prediction.
Limitations and Future Work
The authors acknowledge the following limitations:
- Computational Load: The processing of point clouds and neural network inference, while manageable in their experiments, could pose a challenge for real-time applications on very large surfaces or with faster cycle times.
- Robustness: The system's performance was validated under specific production conditions. Further investigation is needed to assess its robustness across a wider range of concrete setting states and environmental factors.
Future work suggested includes:
- Enhancing computational efficiency for real-time performance.
- Improving robustness under variable operational conditions.
- Advancing system intelligence through dynamic obstacle avoidance and multi-robot collaborative strategies.
Summary updated