This paper maps the real system X of the manipulator(end-effector) to a simpler spring mass types control system Y.

• This transformation is given by a learnable invertible flow network.
• The optimal control $u_y$ is found from the equation an we can get the optimal control for the manipulator state $u_x$ using the inverse jacobian matrix (using the concept of virtual work)
• They also show safety guarantees in their framework

In Normalizing Flows, complex probability distributions are models are constructed by learning a sequence of convertible and differential neural network transformations that map the simple prior distribution to the required complex ones.

Training Details