机器人自动设计初探讨：基于黑盒函数优化实现

Robotics is a research topic among the most funded around the world and robots have the potential to largely reduce human labors. The design of robots is, on the other hand mental labor intensive. It is surprising that automatic robot design, as an important research topic, remains unexplored till today. In this paper, a automatic design framework is proposed, where the robotics design problem is model as a black-box function optimization process. Given robots parts and a specific purpose that the robot is designed to accomplish, the framework tries to auto-assemble the parts in a way that is best for the purpose. The framework is a simulation and optimization interactive process, where a physics engine or even real world simulations are employed to score a assembling setting according to its capability to accomplished the given task, and a black-box optimization method is employed to optimize the score. Bayesian hyper-parameter optimization is used to carry out the optimizations, for its advantage of requiring less sampling than other methods. Experiments show that the proposed method is able to find acceptable designs of a two joint system within only 5 simulations.