Autonomous and remotely operated underwater vehicles allow us to reach places which have previously been inaccessible and perform complex repair, exploration and analysis tasks. As their navigation is not infallible, they may cause severe damage to themselves and their often fragile surroundings, such as flooded caves, coral reefs, or even accompanying divers in case of a collision. In this study, we used a shallow neural network, consisting of interlinking PID controllers, and trained by a genetic algorithm, to control a biologically inspired AUV with a soft and compliant exoskeleton. Such a compliant structure is a versatile and passive solution which reduces the accelerations induced by collisions to 56% of the original mean value acting upon the system, thus, notably reducing the stress on its components and resulting reaction forces on its surroundings. The segmented structure of this spherical exoskeleton protects the encased system without limiting the use of cameras, sensors or manipulators.

自主远程操作的水下航行器使我们能够到达以前无法到达的地方，并执行复杂的修复、勘探和分析任务。由于它们的导航并非万无一失，因此它们可能会对自身及其脆弱的周围环境造成严重损害，例如被淹没的洞穴、珊瑚礁，甚至在发生碰撞时陪同的潜水员。在这项研究中，我们使用了由互连 PID 控制器组成的浅层神经网络，并通过遗传算法进行训练，来控制具有柔软且柔顺的外骨骼的仿生 AUV。这种顺从结构是一种通用的被动解决方案，可将碰撞引起的加速度降低至作用在系统上的原始平均值的 56%，从而显着降低其组件上的应力以及对其周围环境产生的反作用力。这种球形外骨骼的分段结构可以保护封闭系统，而不限制相机、传感器或机械手的使用。