仿人机器人全方位步行参数的代理模型优化

Surrogate Model Based Optimization for Omni-directional Walking Parameters of Humanoids

仿人机器人的全方位步行参数对其行走稳定性、灵活性、快速性具有较大影响,然而物理机器人与描述其前后步幅连接约束的简化动力学模型间的数学关系却难于建立,因而难于获得优化目标表达式和相应的优化方法.本文从步幅跟随规划算法中提取出7个关键影响参数,并将标准实验工况下的步幅跟随性能指定为优化目标,从而将问题转化为一个黑盒优化过程.基于动力学仿真建立Kriging代理模型,通过Latin超立方初始实验和EGO(effective global optimization)迭代建模优化求解该问题.动力学仿真结果表明,在较少的实验代价下,该方法实现了全方位步行参数的优化,该方法能够实现步行速度和步幅跟随能力的综合提升.

Omni-directional walking parameters of a humanoid robot have great impacts on its walking stability, flexibility and rapidity. However, it is difficult to establish the mathematical relationship between the physical robot and the simplified dynamics model which describes its inter-pace connectivity constraints. Therefore, it is hard to obtain the expression of the optimization objective as well as the optimization method. In this paper, seven key effective parameters are extracted from the stride tracking planning algorithm, and the tracking quality of a standard walking experiment is designated as the optimization objective. The problem is thus transferred into a black-box optimization process. Kriging surrogate model is established by dynamics simulation, and the initial experiments designed with Latin Hypercube as well as the EGO（effective global optimization） iterative modeling and optimization are adopted to solve this problem. Dynamics simulation results show that the omni-directional walking parameters are optimized via the proposed approach with quite a little experiment cost. The approach is also able to promote remarkably the walking speed and stride tracking ability simultaneously.