DescriptionThis work aims to develop a framework toward a machine learning-based robotic design algorithm, which employs Compositional Pattern Producing Networks plus NeuroEvolution of Augmenting Topologies (CPPN + NEAT) to optimize the design of mechanical components in a robot. Unlike conventional optimization algorithms, this algorithm can consider the impact of each pixel in a part design on the robot instead of using pre-determined design variables. The algorithm includes four steps. First, generate the geometry of a component in a robot by activating its CPPN networks. Second, employ multiple rules to determine if the geometry meets the requirements of the robot. Third, test the robot with the designed component in the robot simulator Gazebo. Finally, employ the NEAT algorithm to evolve the CPPN networks according to the simulation results. This algorithm has the potential to optimize the robotic design or the mechanical design of a structure in the future.
This work focused on designing a fin-based amphibious robot, accompanied by numerical analysis of a pitching paddle, dynamic modeling of the robot, and machine learning algorithms. The objective of the case study is to optimize the robot’s body. We employ finite element methods to study the hydrodynamic forces of pitching paddles in 2D, then apply the forces to the robot in the simulator. We also create look-up tables to approximate the hydrodynamic forces of the robot’s body with different shapes. To generate a practical design, we use multiple rules to restrict the geometry of the body. A case study shows the detailed procedure of the optimization and proves the feasibility of the algorithm.