基于神经网络的四足机器人 SLIP 模型运动控制

根据四足哺乳动物形态设计的四足机器人，能够适应错综复杂的地形，且具有较强的运动属性。 而对于四足机器人而言，运动的控制十分重要，所以对四足机器人运动过程中的跳跃控制问题进行了研究。 首先，采用弹簧负载倒立摆模型（SLIP）对四足机器人结构进行简化处理，建立了简化模型的动力学方程， 并分析了模型的运动过程以及着陆相与腾空相的转换条件。其次，在仿真平台中建立了 SLIP 动力学模型， 通过动力学仿真得到了一份仿真样本数据，并利用这份样本数据训练了一个神经网络，其中样本考虑了四足 机器人在与地面接触过程中由于碰撞和阻尼所消耗的能量。最后，在仿真平台中进行验证，给定模型的初始 高度和水平速度，通过神经网络计算出合适的着陆角，得到期望的水平末速度和弹跳高度。实验结果表明， 基于神经网络的方法可以较精确地实现对 SLIP 模型运动的控制。

Motion Control of Quadruped Robot SLIP Model Based on the Neural Network

The quadruped robot, which is designed according to the morphology of quadruped mammal, can adapt to the complex terrain with its superior movement property. In view of the importance of motion control required for the quadruped robot, a research has been made of the jump control issue in the movement process of quadruped robots. Firstly, a spring-loaded inverted pendulum model (SLIP) is used to simplify the structure of the quadruped robot, thus establishing the dynamics equation of the simplified model, followed by an analysis of the motion process of the model and the conversion conditions between landing and airborne phases. Secondly, the SLIP dynamics model is to be established in the simulation platform, and the simulation sample data is to be obtained through the dynamics simulation, with a neural network trained by using the sample data, in which the energy consumed by collision and damping in the contact process between the quadruped robot and the ground is to be taken into account. Finally, given the initial height and horizontal velocity of the model, the appropriate landing angle is calculated through the neural network, thus obtaining the expected final horizontal velocity and the bounce height as well. The experimental results show that the method based on the neural network can improve the control accuracy of the SLIP model.