面向协作机器人的零力控制与碰撞检测方法研究

为了解决协作机器人柔顺交互控制问题，本文对机器人的零力控制和碰撞检测方法进行了深入研究。首先将逆运动学问题转化为（Newton-MP）广义逆和牛顿下山法的迭代求解问题。其次，针对协作机器人的零力控制问题，建立了基于速度三次摩擦力模型的完全动力学方程。对摩擦力模型进行遗传算法多参数辨识。再次，提出了基于One-Class卷积神经网络的碰撞检测方法，构建了无碰撞数据集，解决了传统碰撞检测方法建模不准确的问题。最后，通过实验证明，本文提出的Newton-MP优化方法具有良好的性能，绝对误差达到0.00013mm。与理想摩擦力模型进行对比，采用基于速度的三次摩擦力模型拟合出的摩擦力能够更好适用于零力控制。将外力矩观测器与One-Class卷积神经网络碰撞检测进行优缺点分析，可以证明One-Class卷积神经网络可以在不依靠模型的情况下，准确地检测机器人的异常碰撞。

Research on Zero-force Control and Collision Detection for Cooperative Robots

In order to solve the problem of compliant interactive control of cooperative robots, this paper studies the zero-force control and collision detection methods of robots. First, it transformed the inverse kinematics problem into an iterative solution problem of the Newton-MP method. Secondly, to solve the zero-force control problem of cooperative robots, a complete dynamic equation based on the velocity cubic friction model is established. A genetic algorithm identified the friction model with multiple parameters. Thirdly, the paper proposed a collision detection method based on a One-Class convolution neural network. It constructed the un-collision dataset to solve the problem of inaccurate modeling of traditional collision detection methods. Finally, the experiment proves that the Newton-MP method performs well, and the absolute error reaches 0.00013 mm. Compared with the ideal friction model, the friction fitted by the cubic friction model based on velocity is more suitable for zero force control. By analyzing the collision detection method of the external moment observer and the One-Class convolution neural network, it can be proved that the One-Class convolution neural network can accurately detect the robot''s abnormal collision without depending on the model.