Per-Finger Prosthetic Grasp Planning Using Object-Aligned Bounding Box Representation and VLM-Driven Object Selection

Recent progress in prosthetic manipulation highlights the need for perception-driven control strategies that can adapt to diverse objects and user intent. This work presents a modular vision-guided grasping pipeline that integrates VLM-based object identification, orientation-aligned geometric modeling, and per-finger grasp planning for dexterous prosthetic hands. A Vision–Language Model (VLM) identifies the target object and activates the grasping pipeline only when recognition is confident, supporting intent-aware operation. From the segmented point cloud, an object-aligned bounding box (OBB) is constructed to provide a compact, orientation-aware representation of the object’s global extents, enabling more accurate distance and collision queries than axis-aligned boxes. Using this representation, the system evaluates candidate fingertip trajectories and selects contact poses for each finger independently, followed by Damped Least Squares inverse kinematics for joint-level execution. Preliminary experiments on a limited set of representative objects using the Linker Hand O7 demonstrate that the proposed pipeline achieves consistent grasp execution and exhibits promising real-time performance within controlled scenarios. In simulation, the proposed pipeline achieved a maximum segmentation accuracy of 93.4%, while hardware experiments on the Linker Hand O7 achieved 93.2% segmentation accuracy, confirming stable grasp execution across representative objects. While the evaluation is not yet comprehensive, the results indicate that combining semantic object identification with lightweight geometric reasoning can support efficient and adaptable grasp generation suitable for future prosthetic applications.