Variance-Reduced MPPI: Workshop Version

Fabian Schramm, Franki NGUIMATSIA TIOFACK, Nicolas Perrin-Gilbert, Marc Toussaint, Justin Carpentier

Keywords: sampling-based optimization, robotics

Abstract: Sampling-based controllers like Model Predictive Path Integral (MPPI) offer substantial flexibility but suffer from high variance and low sample efficiency. To address this, we introduce a variance-reduced MPPI framework that decomposes the objective into an approximate model and a residual term. By adopting a quadratic approximation, we derive a closed-form, model-guided prior that concentrates samples in informative regions. Crucially, this framework is agnostic to the source of geometric information, accommodating exact derivatives, structural approximations (e.g., Gauss-Newton), or gradient-free randomized smoothing. We validate our approach on optimization benchmarks, underactuated cart-pole control, and contact-rich manipulation and show that it achieves faster convergence in low-sample regimes compared to standard MPPI.

Submission Number: 41