Linkage structures are primitive systems that are useful in industrial fields.However, their sparse characteristics make it difficult to apply topological optimization techniques, except in the field of statics.We proposed a continuous feature field for the linkage structure via a generative neural network, which allows for topology optimization in the dynamic kinematics field. In order to examine this, a conditional generative adversarial network (cGAN) was employed, and it was connected with a differentiable multi-body simulator. It was demonstrated that the rigid-body multi-pendulum systems transitioned in a stepwise manner, in accordance with the gradient of the objective function.Consequently, it is substantiated that the number of links, their connecting topology, and the scale of the linkage bar can be updated along the objective.This approach expands the benefits of differentiable simulators, such as policy-gradient-based behavior optimization and parameter tuning, to include topology optimization.
YONEZAWA et al. (Wed,) studied this question.