New Paper Shows MPPI Control Is Actually Expectation-Maximization Algorithm
Researchers reinterpret Model Predictive Path Integral as EM, enabling broader applications.
Jiarui Wang, Sina Sharifi, and Mahyar Fazlyab show that Model Predictive Path Integral (MPPI) control is a special case of the Expectation-Maximization (EM) algorithm applied to a probabilistic inference formulation of optimal control. Their generalized EM-MPPI framework extends MPPI beyond Gaussian parameterization, analyzes convergence behavior, and characterizes local convergence rates. For exponential-family distributions, they prove a sufficient increase property when the log-partition function is strongly convex, and specializing to Gaussian MPPI yields explicit global and local convergence characterizations.
- MPPI control is reinterpreted as EM algorithm for probabilistic optimal control, providing a rigorous theoretical foundation.
- New generalized EM-MPPI framework works with non-Gaussian distributions, such as exponential families, for more flexible exploration.
- Local convergence rate is characterized by covariance of posterior trajectory distribution, with explicit global convergence for Gaussian MPPI.
Why It Matters
Bridges sampling-based control and statistical inference, enabling more flexible and theoretically grounded robot controllers.