Learning to Correspond Dynamical Systems
share
Correspondence across dynamical systems can lend us better tools for learning optimal control policies in new systems. We present a fully data-driven approach for putting multiple dynamical systems into correspondence. We utilize symmetric nearest neighbor distances in our loss function to encourage states of different sources to reside in close proximity in a shared latent state space. We also construct and enforce a dynamics in the latent space, which helps avoid temporal ambiguity of state projections. We demonstrate the effectiveness of our approach by putting a 2D walker character into correspondence with a PD-controlled pendulum and with a 2D walker character with a different morphology.
READ FULL TEXT