Neural Stochastic Contraction Metrics for Learning-based Robust Control and Estimation

  • 2020-11-19 17:39:43
  • Hiroyasu Tsukamoto, Soon-Jo Chung, Jean-Jacques E. Slotine
  • 0

Abstract

We present Neural Stochastic Contraction Metrics (NSCM), a new designframework for provably-stable robust control and estimation for a class ofstochastic nonlinear systems. It uses a spectrally-normalized deep neuralnetwork to construct a contraction metric, sampled via simplified convexoptimization in the stochastic setting. Spectral normalization constrains thestate-derivatives of the metric to be Lipschitz continuous, thereby ensuringexponential boundedness of the mean squared distance of system trajectoriesunder stochastic disturbances. The NSCM framework allows autonomous agents toapproximate optimal stable control and estimation policies in real-time, andoutperforms existing nonlinear control and estimation techniques including thestate-dependent Riccati equation, iterative LQR, EKF, and the deterministicneural contraction metric, as illustrated in simulation results.

 

Quick Read (beta)

loading the full paper ...