I recently started learning about DQN agents and how they can be applied to simple robot arms, as well as seeing a lot of use in virtual settings. I also saw a paper from Cornell on making its application to multi link robots much smoother to train.
My question is, could this approach ever fully replace the need for building kinematic models / denavit hartenberg parameters and the like? Pushing into territory I am even hazier on, would it also solve the dynamic aspect (I believe jacobians are needed here?)?