Bio-Inspired High-Precision Vision-Based Hand-Eye Calibration via an Integrated Target

Humans perform precision manipulation by continuously watching their own hands: vision delivers high-precision localization that compensates for the low fidelity of proprioception, and the internal eye–hand mapping is calibrated and refined through visual feedback. Inspired by this biological strategy, this paper proposes a bio-inspired high-precision vision-based hand–eye calibration method using an Integrated Target. A global RGB-D camera fixed above the workspace, playing the role of the human eye, directly measures the absolute pose of the target rigidly integrated with the hand–eye camera, replacing the low-precision controller-reported end-effector poses (analogous to proprioception) on one side of the calibration equation, while the hand–eye camera observes a stationary calibration board to provide the other. Four techniques support this framework: (i) a rigid-body Integrated Target design eliminating the geometric uncertainty of separate mounting; (ii) a multi-modal conditional random field fusing kinematic priors, color appearance, and depth geometry for high-fidelity target point cloud segmentation; (iii) a three-stage cascaded registration strategy combining kinematic priors with association graph matching, overcoming local optima under self-occlusion and weak textures; and (iv) a registration-residual-aware weighted optimization that maps the end-to-end registration root-mean-square error (RMSE) to relative confidence weights via a Gaussian kernel on the SE(3) manifold. Experiments on a Franka Emika Panda (absolute positioning error ±2 mm) demonstrate a segmentation intersection over union (IoU) of 0.9760. Stage 2 coarse registration achieves a mean RMSE of 0.3887 mm with a mean runtime of 0.0961 s per pose, whereas the complete three-stage registration requires 0.49 s per pose. The calibrated measurement system achieves a mean absolute sphere-center distance error of 0.3068 mm. Compared with the kinematics-based baseline (8.92 mm), analogous to blindfolded manipulation, the measurement error is reduced by more than an order of magnitude. In a separate target-level comparison with an ArUco marker target, the Integrated Target achieves valid-observation coverage of 94.7% versus 50.7% over the sampled viewpoints and a mean absolute displacement error of 0.12 mm versus 1.23 mm under controlled 20 mm translations. These results characterize target observability and translational localization, rather than the accuracy of the calibrated hand–eye transformation.