Bimbraw Keshav, Fox Elizabeth, Weinberg Gil, Hammond Frank L
Annu Int Conf IEEE Eng Med Biol Soc. 2020 Jul;2020:4753-4757. doi: 10.1109/EMBC44109.2020.9176483.
Sonomyography (ultrasound imaging) offers a way of classifying complex muscle activity and configuration, with higher SNR and lower hardware requirements than sEMG, using various supervised learning algorithms. The physiological image obtained from an ultrasound probe can be used to train a classification algorithm which can run on real time ultrasound images. The predicted values can then be mapped onto assistive or teleoperated robots. This paper describes the classification of ultrasound information and its subsequent mapping onto a soft robotic gripper as a step toward direct synergy control. Support Vector Classification algorithm has been used to classify ultrasound information into a set of defined states: open, closed, pinch and hook grasps. Once the model was trained with the ultrasound image data, real time input from the forearm was used to predict these states. The final predicted state output then set joint stiffnesses in the soft actuators, changing their interactions or synergies, to obtain the corresponding soft robotic gripper states. Data collection was carried out on five different test subjects for eight trials each. An average accuracy percentage of 93% was obtained averaged over all data. This real-time ultrasound-based control of a soft robotic gripper constitutes a promising step toward intuitive and robust biosignal-based control methods for robots.
超声成像技术提供了一种对复杂肌肉活动和形态进行分类的方法，与表面肌电图相比，它具有更高的信噪比和更低的硬件要求，并且可以使用各种监督学习算法。从超声探头获得的生理图像可用于训练一种能够在实时超声图像上运行的分类算法。然后，预测值可以映射到辅助或远程操作机器人上。本文描述了超声信息的分类及其随后映射到软机器人夹爪上的过程，作为迈向直接协同控制的一步。支持向量分类算法已被用于将超声信息分类为一组定义的状态：张开、闭合、捏握和钩状抓握。一旦用超声图像数据训练了模型，就使用来自前臂的实时输入来预测这些状态。最终预测状态输出随后设置软致动器中的关节刚度，改变它们的相互作用或协同作用，以获得相应的软机器人夹爪状态。对五名不同的测试对象进行了数据收集，每人进行八次试验。所有数据的平均准确率为93%。这种基于实时超声的软机器人夹爪控制是迈向直观且强大的基于生物信号的机器人控制方法的有希望的一步。
