Sensorless Contact Force Estimation on Raven II Surgical Robot Platform

Abstract

Haptic feedback in surgical robotic arms has long been a challenging area due to factors such as the need for sterilization, which complicates the mounting of force sensors on the end-effector, and the complex, nonlinear nature of cable-driven systems. In this work, I present a data collection procedure that enables contact force estimation without direct sensor measurements, using a learning-based approach. Additionally, a 6-degree-of-freedom (DOF) cable-driven haptic device was developed to apply known contact forces over a wide workspace. The device includes multiple direct-current (DC) motors and shear beam load cells, and a real-time feedback control system implemented using the Zephyr Real-Time Operating System (RTOS). A high-level controller synthesizes the desired contact forces. The device was tested on a Raven II Surgical Robot Arm, demonstrating its capabilityin applying precise contact forces, though certain limitations in the setup were identified. A one-directional contact force estimation experiment was conducted with a random force applied along the Y-axis, while the Raven-II followed a random trajectory. Three neural network models—MLP, LSTM, and GRU—were trained using the collected data, achieving mean absolute errors (MAE) of 0.5859, 0.6664, and 0.7486 N, respectively, within a force range of ±4.5 N. Further analysis revealed a correlation between joint velocities and estimation error, providing insights for designing more efficient training trajectories.