Grip force measurement enables better control performance for robot-assisted minimally invasive surgery (RMIS). A deep-learning-based method is proposed to measure the instrument grip force without mounting additional sensors in this article. First, the training trajectory and input data frame are studied. Seven data are derived from the original sensor data, and a binary butterfly optimization algorithm with opposite-based learning (bBOA-OBL) is conducted to form the suitable input data frame. Based on the data frame, a novel convolutional network with an attention mechanism and feedforward of current (CAM-FoC) is proposed to calculate the grip force. The results of the ablation study tell that the master-slave trajectory has the highest accuracy, the optimized input data frame can reduce the error by 17%, and each component of CAM-FoC can enhance the measurement accuracy. Experiments and comparisons are also carried out. The root mean squared error (RMSE) in the experiment is only 0.1233 N and is lower than four other popular methods. In addition, the average computation time is around 2 ms on different platforms. The method manifests higher measurement accuracy than the state-of-the-art and is of acceptable computational complexity. The technology would potentially achieve grip force feedback in clinical usage, which can significantly improve surgery performance. The method could also be used by other robots composed of the cable-pulley mechanism to acquire external forces.
19 March 2020 in IEEE Transactions on Cognitive and Developmental Systems
Obtaining grip force in robot-assisted minimally invasive surgery (RMIS) has always been a crucial research topic. A novel grip force cognition scheme is proposed in this paper. We utilize the dynamic analysis of the cable-driven system to conduct feature engineering and add prior knowledge to Gaussian Process Regression (GPR). Specifically, feature and GPR model candidates are acquired by analyzing the dynamic characteristics first. After that, a wrapper searching method combined with Exhaustive Search (ES) and Sequential Backward Searching (SBS) is proposed to determine the features and GPR model. The hyper-parameters are optimized by Differential Evolution (DE) through minimizing Negative Log Marginal Likelihood (NLML). Besides, the training set is extended by including several different objects to avoid the algorithm “learning” the object property. Our method is verified on a cable-driven platform to ensure the errors are all derived from the algorithm rather than cable tension loss. Our method reduces the errors on our testing set by about 64% and outperforms several popular model based and learning based methods. This work illustrates the significance to integrate the system’s characteristics when introducing machine learning techniques into robot systems.
3 March 2023 in IEEE Robotics and Automation Letters
Learning-based grip force measurement methods in robot-assisted minimally invasive surgery (RAMIS) outperforms the traditional model-based methods and avoids the application issues of sensor-based approaches. However, few studies have investigated the problem of grip force measurement in mass-produced surgical instruments. This letter takes the difference in motion hysteresis and mechanism friction among mass-produced surgical instruments into account, and proposes a novel learning-based method ACAM-FoC. ACAM-FoC is an augmentation of CAM-FoC, which is a high accuracy lightweight network proposed in our preceding research. Loss function monotonicity limitation is introduced to alleviate the negative effect of unbalance training data. Offiline experiments are conducted to verify the effectiveness and the advantages over other two existing methods. Online experiments on the self-developed surgical robot system are conducted to preliminarily realize and verify the grip force measurement and feedback in real time. The results support the importance of grip force feedback in leader-follower operation tasks.