Efficient data-driven joint-level calibration of cable-driven surgical robots

Abstract

Accurate joint position estimation is crucial for the control of cable-driven laparoscopic surgical robots like the RAVEN-II. However, any slack and stretch in the cable can lead to errors in kinematic estimation, complicating precise control. This work proposes an efficient data-driven calibration method, requiring no additional sensors post-training. The calibration takes 8–21 min and maintains high accuracy during a 6-hour heavily loaded operating. The Deep Neural Network (DNN) model reduces errors by 76%, achieving accuracy of 0.104∘, 0.120∘, and 0.118 mm for joints 1, 2, and 3, respectively. Compared to end-to-end models, the DNN achieves better accuracy and faster convergence by correcting original inaccurate joint positions. Additionally, a linear regression model offers 160 times faster inference speed than the DNN, suitable for RAVEN’s 1000 Hz control loop, with slight compromises in accuracy. This approach should significantly enhance the accuracy of similar cable-driven robots.