Model-Free Design Optimization of a Hopping Robot and Its Comparison With a Human Designer

Abstract

When developing a robot, design iterations in the physical world are necessary, even though they are often costly and not systematic. Here, we present an automated iterative design process without using modeling or simulation, which we refer to as “model-free design optimization” based on Bayesian optimization. This letter particularly focuses on the cooptimization of morphology and controller, by using a mechanism to balance parameter specific costs (i.e., morphology samplings are more expensive than control ones) for effective and efficient design optimization processes. A hopping robot was employed for a feasibility analysis of the proposed optimization method, in which minimalistic two-dimensional and four-dimensional design optimization experiments were performed in real life. The results show that the proposed approach is capable of improving both of the robot design problems within a defined time limit. The method is also compared to optimization performances of a human designer under the same conditions. The automated method has advantage in finding the best solution more quickly in the four-dimensional search space, while the human optimization performs better in the two-dimensional case.