Ablation Study on Features in Learning-Based Joint Calibration of Cable-Driven Surgical Robots

Abstract

Surgical robots equipped with cable-driven mechanisms have flexible, light, and compact arms and tools. However, cable slack, stretch, and gear backlash introduce unavoidable errors from motor positions to joint positions and the end-effector pose. This paper presents a learning-based joint position calibration method for the RAVEN-II surgical robot, employing deep neural networks and gated recurrent units. Compared to fixed offset compensation, the learning-based calibrations reduce the joint position errors by over 62.4% (unloaded) and 54.8% (loaded). Furthermore, removal and inaccurate ablation studies on input features identify that raw joint positions and motor torques are the most important model inputs for calibration accuracy. These studies also reveal that the models are capable of inferring joint positions from the end-effector pose and prioritize the direction of motor torques over their amplitude. When guided appropriately, the models can also compensate for encoder value inconsistencies occurring with robot re-homings. By excluding the unnecessary input features, lightweight models are developed and achieve better performance and efficiency simultaneously, reducing the training time on the CPU to 2.5 minutes. All data and code are open-source at https://github.com/uw-biorobotics/RAVEN-2-Feature-Ablation Note to Practitioners—This paper presents a data-driven neural network/AI based calibration method for accurate joint position estimation on a cable-driven research surgical robot (RAVEN-II). We aim to understand our learning calibration by studying the importance of each input feature using two ablation methods. We selectively eliminate (removal ablation) or add distorting noise to (inaccurate ablation) input features one at a time, and retrain the calibration model. Then, the increase in error suggests the importance of the target input feature. By excluding unnecessary input features and thus reducing the input dimension, the size of the machine learning model can be reduced without losing accuracy. The simpler model can be trained and inferred efficiently without GPU acceleration. All data and code used in this paper, including robot control, data collection, and model training, are available online.