Boundary-Constrained Vision and Spectral Point-Cloud Filtering for 6-DoF Robotic Grasping of Hollow-Frame Porous Structures

Abstract Automated sorting and precision robotic manipulation of lightweight hollow- frame workpieces, such as lattice-style micro-unmanned aerial vehicle (micro- UAV) frames and topology-optimized industrial components, present acute perceptual bottlenecks for standard 3D visual sensing. Active infrared struc- tured light efortlessly traverses internal apertures and void cavities, interleaving upper surface contact points with underlying support table clutter and severely destabilizing 6-DoF grasp pose estimation. To resolve this geometric contam- ination without relying on CAD models or prior 3D templates, this paper proposes an integrated boundary-constrained perception and geometry-aware grasping pipeline tailored for porous and hollow-frame structures. The proposed architecture establishes an external-perimeter segmentation strategy utilizing a customized YOLOv8-seg network, explicitly masking out interior void regions at the visual source. The segmented depth cluster is back-projected into 3D camera coordinates and processed through a nearest-20% depth-quantile filter ( Q0 . 20 ) to decouple the top planar contact rim from vertical sidewalls. Local surface nor- mal vector extraction is formulated via Singular Value Decomposition (SVD) of the spatial scatter covariance matrix, accompanied by an analytical planarity spectral-ratio gate ( κ = λ 1 /λ3 ≥ 15 . 0 ) to reject non-planar degeneracies. In- plane rotational ambiguity is resolved by projecting the 2D contour axis onto the tangent plane, synthesizing robust grasp orientations on SO(3) via Lie algebra matrix exponential mapping. Perception and motion execution are coordinated across a distributed mechatronic testbed linking an industrial 6-DoF Aubo-i5 collaborative manipulator, an eye-in-hand RealSense D435i sensor, and an SMC pneumatic parallel gripper over a lightweight asynchronous JSON-RPC protocol. Experimental evaluations demonstrate a boundary segmentation mask [email protected] of 0.995 across 106 annotated frames, alongside 68 physical workspace grasping trials yielding an empirical position jitter of 0.167 mm, an orientation jitter of 0 . 249◦ , and an overall geometric convergence rate of 98.5% (67/68 trials), verify- ing the system’s operational viability and numerical robustness under real-world laboratory conditions.