Monocular Depth-Based Object Localization for Collaborative Robotic Grasping

Abstract

This paper investigates the application of monocular depth estimation for 3D object localization in collaborative robotic manipulation tasks. We propose a method for converting relative depth maps obtained using the Depth Anything V2 Small model into metric depth maps by employing small 5×5 cells ArUco markers with a physical size of 3×3 cm, which do not obstruct the operator or reduce the available workspace. Pre-calibration on 10–30 images achieves a diameter measurement accuracy of mean absolute error (MAE) of 0.307-0.671 cm (relative error 3.6-7.9%) and height measurement accuracy of a 1.73–2.13 cm MAE (relative error 14.6-16.7%). The full processing cycle for a single frame takes 1.19 seconds on a mid‑range laptop without a dedicated GPU, which confirms the method’s suitability for operation under modest frame rate requirements. The proposed approach is applicable in collaborative scenarios with low‑speed object motion and can be recommended for robots equipped only with an RGB camera, provided that pre‑calibration is performed. The method is characterized by its ease of implementation, low cost, and the absence of a need for specialized hardware, making it attractive for a wide range of human-robot interaction tasks in constrained spaces.