How can a robot drill perfectly on curved surfaces in space?

In the era of Industry 5.0, robots are no longer limited to simple repetitive tasks: they must understand complex shapes and adapt in real time. A research team has developed a robotic drilling system that can handle aerospace components with extreme curvature.

The core of the technology is an adaptive surface‑normal estimation using PCA (Principal Component Analysis) on point clouds. A 660 nm laser projector and a 1.3 MP industrial camera, fixed at a 30° angle, project and capture structured‑light fringes. Triangulation reconstructs a high‑resolution point cloud. The software dynamically selects local regions, computes normal vectors, and corrects the drill path before drilling starts.

Tests on a 400 mm cylinder show that, using points within a 100 mm radius, normal‑estimation error stays below 0.2°, while hole position accuracy is within 0.25 mm, even on rough or thin‑walled surfaces. This far outperforms manual methods, cutting waste and speeding up the assembly of aerospace structures.

The outcome is higher reliability and productivity for companies building satellites, aircraft, or drones, where every hole must meet tight tolerances to ensure component safety and longevity.

The proposed method combines computer vision and multivariate statistics. Structured light generated by a 660 nm laser projector creates a fringe pattern on the part's surface. When the camera captures these fringes, triangulation converts image coordinates into 3‑D coordinates, producing a point cloud with sub‑millimetric resolution.

PCA is applied to local subsets of the cloud. Practically, for each selected region, the algorithm computes the covariance matrix of the points and extracts its eigenvectors. The eigenvector associated with the smallest eigenvalue represents the local normal direction, as it indicates the axis of minimal variance. Adaptive region selection is crucial: in areas with high curvature or roughness, the search radius is reduced to maintain data coherence, while on smoother surfaces it can be expanded to reduce noise.

Once the normal is estimated, the system calculates the angular correction needed for the drill, ensuring the drilling axis is perpendicular to the actual surface, not the idealized one. After drilling, a new scan verifies the hole position by comparing the post‑process point cloud with the CAD model. The average position error (≤0.25 mm) and normal error (≤0.2°) remained consistent across curvature radii (50 mm‑300 mm) and roughness levels (Ra = 0.8 µm‑2.5 µm).

From an industrial perspective, integrating this system into robotic production lines drastically reduces reliance on skilled operators for manual calibration, cutting set‑up times and defect costs. Moreover, real‑time operation enables predictive machining strategies, where the robot can anticipate deformations caused by cutting forces or thermal variations, adjusting the trajectory on the fly. This approach marks a significant step toward smart factories, where micrometric precision is guaranteed by autonomous algorithms, boosting the competitiveness of the aerospace sector and other high‑tech industries.