DNN-Based Force Estimation in Hyper-Redundant Manipulators

Abstract

Hyper-redundant manipulators (HRMs) have been applied to many robotic surgery platforms because they can be bent in various directions as they consist of many rigid segments and are operated in a cable-driven manner, allowing remote placement of actuators. However, due to the long kinematic chain and wire stretching when subjected to force, HRMs have disadvantages, including inaccurate positioning and difficulty in building physics-based models. These constraints pose challenges in realizing force feedback using HRMs. In this paper, we present a deep neural network (DNN) based estimation system for the external force exerted on the distal end of the HRMs using tensions and fixed positions of the cables as input data, which can be measured outside the human body. To generate a training dataset for DNN, a physics-based HRM model considering external forces, cable tensions, cable friction, and cable extension was developed. We also proposed the iterative optimization algorithm for the optimal configuration of the HRM. The training dataset was produced through simulation using the physics-based HRM model, and then the DNN-based force estimation model was trained. The HRM prototype was fabricated to verify the reliability of the training dataset and evaluate the performance of the DNN-based force estimation model. The estimated magnitude and angle errors of the external force were 4.8% and 10.4% on average, respectively. This demonstrates the feasibility of measuring the external force without force/torque sensors, which can help realize force feedback in HRMs.