Stochastic Stability of Nonlinear MPPI via Contraction Theory and Control Lyapunov Functions

What the math gives to ML

The paper provides a usable stability-inheritance pattern for stochastic, sampling-based policies: approximate a contractive reference controller, decompose approximation error into temperature bias and Monte Carlo error, and enforce a small-gain condition on the local error gain. This can transfer to neural world-model controllers by using a learned policy as the nominal controller and an MPPI-like rollout head as a stochastic correction mechanism. The key engineering asset is the explicit allocation of approximation error between temperature, sample count, and contraction margin. A practical implementation can adapt rollout count and temperature online using an estimated local gain, then test whether closed-loop instability decreases at a fixed computation budget.