Pose prediction and compensation method for parallel robot based on physics-informed transfer learning

• Physics-informed transfer learning achieves high-precision pose prediction for parallel robots with limited experimental data. • SHAP analysis quantifies slider-pose directional dependencies, guiding coordinated compensation strategies. • Comprehensive error index defines transfer learning applicability across three sample-size regimes. • Experiments demonstrate 92.5 % position and 72.7 % orientation error reduction, outperforming traditional calibration by significant margins. Parallel robot end-effector pose accuracy directly determines processing quality and system stability, making high-precision pose compensation essential. However, traditional model-driven calibration methods require extensive per-robot measurements to establish parametric error models, limiting their transferability and scalability, while data-driven neural network methods require extensive labeled data that is costly to acquire in industrial settings. This study proposes a novel pose prediction and compensation method based on physics-informed transfer learning (TL) strategy that leverages kinematic and dynamic prior knowledge from theoretical models to enhance pose prediction and compensation accuracy under limited experimental data. First, a physics-informed transfer learning model integrating kinematic and dynamic priors is proposed. Controlled experiments comparing single kinematic model versus kinematic and dynamic model source domains confirm that incorporating dynamic modeling to capture load-dependent deformations significantly enhances knowledge transfer effectiveness. Second, we employ the SHapley Additive exPlanations (SHAP) method to quantify and reveal that the moving platform pose and slider displacement exhibit significant directional dependency. Third, we propose a comprehensive error index to determine the applicability boundary of the TL model. It is found that, in the small sample regime (50–120 samples, Stage B), the TL model demonstrates significantly superior prediction performance compared with other models, and it is the only method achieving effective pose compensation. Compared with the Modified Denavit-Hartenberg (MD-H) method, the TL model reduces the maximum absolute error of position and orientation by 92.5% ( y direction) and 62.2% ( β direction), respectively, which fully verifies its effectiveness, portability, and engineering application value under complex working conditions.