2019-05-18
[1] describes a relative easy to understand task. The idea is, that a robot arm who looks like the snake robot of Marvin Minsky [2] has an inverse kinematics. The kinematic chain is forming the forward model and this is learned by a deep neural network. In contrast, to a normal industrial robot, the frames of the chain are elastic, which means that the construction is complete out of control and the neural network has to learn an approximation of the real physical model. In the literature this concept is known as model-predictive control and it's a powerful technique for control a system. The advantage of this paper is, that the problem isn't very complicated and the described workflow make sense. In the headline of the journal it's written that the journal understands themself as “original research”, i would it a well written educational article about a topic which was introduced in the past.
[1] Model-Based Control of Soft Actuators Using Learned Non-linear Discrete-Time Models https://www.frontiersin.org/articles/10.3389/frobt.2019.00022/full
[2] Moran, Michael E. "Evolution of robotic arms." Journal of robotic surgery 1.2 (2007): 103-111.