Nimble, developed at Stanford, is an analytically differentiable fork of the popular DART physics engine. That means everything from fast Jacobians of dynamics, to using a physical timestep as a non-linearity in your PyTorch neural networks!