Monocular vision-constrained robot kinematic calibration

Abstract

Monocular vision-based robot calibration over a wide orientation range is susceptible to planar-feature degeneracy and parameter coupling. This paper presents an integrated calibration method that combines a polyhedral coded target with kinematically constrained bundle adjustment (KBA) in the pixel domain. Noncoplanar features distributed across multiple target surfaces improve perspective-n-point (PnP) observation geometry. Singular value decomposition (SVD) and column-pivoted QR decomposition determine the identifiable dimension and select a stable subset of modified Denavit-Hartenberg (MDH) parameters, while KBA jointly estimates the selected MDH parameters, system extrinsic transformations, and camera intrinsic and distortion parameters. In an equal-point-count ablation study, the median condition number of the PnP reprojection Jacobian decreased from 131.95 to 50.65. Calibration was performed using 64 configurations, and evaluation used 368 held-out configurations. The vision-based flange-position root-mean-square error (RMSE) decreased from 4.742 to 0.827 mm, and the training reprojection RMSE decreased from 3.731 to 0.519 px. Against an external laser tracker (LT) reference, the flange-position RMSE decreased from 2.729 to 2.364 mm. These results show that the method improves observation geometry and out-of-sample consistency over a wide orientation range, while providing only a modest improvement in external position accuracy.