Overcoming Initial Convergence in Multi-objective Evolution of Robot Control and Morphology Using a Two-Phase Approach
Overcoming Initial Convergence in Multi-objective Evolution of Robot Control and Morphology Using a Two-Phase Approach
复制标题
使用两阶段方法克服机器人控制和形态学多目标进化中的初始收敛
DOI：
--
复制
发表时间：
2017
期刊：
影响因子：
--
通讯作者：
K. Glette
中科院分区：
文献类型：
--
作者：
T. Nygaard;Eivind Samuelsen;K. Glette
机器人形态和控制系统的协同进化是机器人设计的一种新方法。然而，搜索空间的增加的大小和强度成为一个挑战，往往导致早期收敛与次优形态控制器组合。此外，从控制器的角度来看，机器人形态的突变往往会导致搜索中的大扰动，从而有效地改变环境。在本文中，我们提出了一个两阶段的方法来解决早期收敛的形态控制器协同进化。在第一阶段中，我们允许形态和控制器同时自由进化，而在第二阶段中，我们在锁定形态的同时重新进化控制器。该方法的可行性在物理仿真中得到了证明，并在三个不同的机器人形态的真实世界的情况下进行了验证。结果表明，通过引入两阶段的方法，搜索产生的解决方案，优于单一的共同进化运行超过10%。
Co-evolution of robot morphologies and control systems is a new and interesting approach for robotic design. However, the increased size and ruggedness of the search space becomes a challenge, often leading to early convergence with sub-optimal morphology-controller combinations. Further, mutations in the robot morphologies tend to cause large perturbations in the search, effectively changing the environment, from the controller’s perspective. In this paper, we present a two-stage approach to tackle the early convergence in morphology-controller co-evolution. In the first phase, we allow free evolution of morphologies and controllers simultaneously, while in the second phase we re-evolve the controllers while locking the morphology. The feasibility of the approach is demonstrated in physics simulations, and later verified on three different real-world instances of the robot morphologies. The results demonstrate that by introducing the two-phase approach, the search produces solutions which outperform the single co-evolutionary run by over 10%.