Research Centre for Medical Robotics and Minimally Invasive Surgical Devices, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, 1068 Xueyuan Avenue, Shenzhen 518055, China; Shenzhen College of Advanced Technology, University of Chinese Academy of Sciences, Shenzhen 518055, China; CAS Key Laboratory for Health Informatics, Shenzhen Institutes of Advanced Technology, Shenzhen 518055, China.
Research Centre for Medical Robotics and Minimally Invasive Surgical Devices, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, 1068 Xueyuan Avenue, Shenzhen 518055, China.
Neural Netw. 2018 Nov;107:34-47. doi: 10.1016/j.neunet.2018.06.018. Epub 2018 Aug 23.
Recently, snake-like robots are proposed to assist experts during medical procedures on internal organs via natural orifices. Despite their well-spelt advantages, applications in radiosurgery is still hindered by absence of suitable designs required for spatial navigations within clustered and confined parts of human body, and inexistence of precise and fast inverse kinematics (IK) models. In this study, a deeply-learnt damped least squares method is proposed for solving IK of spatial snake-like robot. The robot's model consists of several modules, and each module has a pair of serial-links connected with orthogonal twists. For precise control of the robot's end-effector, damped least-squares approach is used to minimize error magnitude in a function modeled over analytical Jacobian of the robot. This is iteratively done until an apt joint vector needed to converge the robot to desired positions is obtained. For fast control and singularity avoidance, a deep network is built for prediction of unique damping factor required for each target point in the robot's workspace. The deep network consists of 11 x 15 array of neurons at the hidden layer, and deeply-learnt with a huge dataset of 877,500 data points generated from workspace of the snake robot. Implementation results for both simulated and actual prototype of an eight-link model of the robot show the effectiveness of the proposed IK method. With error tolerance of 0.01 mm, the proposed method has a very high reachability measure of 91.59% and faster mean execution time of 9.20 (±16.92) ms for convergence. In addition, the method requires an average of 33.02 (±39.60) iterations to solve the IK problem. Hence, approximately 3.6 iterations can be executed in 1 ms. Evaluation against popularly used IK methods shows that the proposed method has very good performance in terms of accuracy and speed, simultaneously.
最近，蛇形机器人被提议通过自然孔道协助专家在内部器官上进行医疗程序。尽管它们具有明显的优势，但在放射外科中的应用仍然受到缺乏在人体密集和受限部位进行空间导航所需的合适设计以及缺乏精确和快速的逆运动学 (IK) 模型的阻碍。在这项研究中，提出了一种深度学习的阻尼最小二乘法来解决空间蛇形机器人的 IK 问题。机器人的模型由几个模块组成，每个模块都有一对与正交扭曲连接的串联连杆。为了精确控制机器人的末端执行器，使用阻尼最小二乘法来最小化模型中经过机器人分析雅克比的函数的误差幅度。这是通过迭代完成的，直到获得将机器人收敛到期望位置所需的合适关节向量。为了快速控制和避免奇点，为机器人工作空间中的每个目标点预测独特阻尼因子构建了一个深度网络。深度网络在隐藏层中由 11 x 15 个神经元组成，并使用 877,500 个数据点的大数据集进行深度学习，这些数据点是从蛇形机器人的工作空间生成的。机器人的八连杆模型的模拟和实际原型的实现结果都表明了所提出的 IK 方法的有效性。在 0.01 毫米的误差容限下，该方法具有非常高的可达性度量，收敛的平均执行时间为 9.20（±16.92）毫秒。此外，该方法解决 IK 问题平均需要 33.02（±39.60）次迭代。因此，大约 3.6 次迭代可以在 1 毫秒内执行。与常用的 IK 方法的评估表明，该方法在准确性和速度方面同时具有非常好的性能。
