Related publications
- Toward Single-Step MPPI via Differentiable Predictive Control2026 IEEE 65th Conference on Decision and Control (CDC), 2026
Overview
MPPI can handle nonlinear control problems through trajectory sampling, but longer horizons increase its computational and sampling requirements. Step-MPPI learns to carry longer-horizon objectives into a single-step lookahead controller.
Method
A neural network predicts the sampling mean and covariance. Cost weights and temperature are learned jointly using the control objective, constraint penalties, and maximum-entropy regularization. Training embeds future consequences in the sampling policy and cost, reducing the work required at deployment.
Evaluation
The paper studies high-dimensional systems and tasks with long control horizons, targeting millisecond-scale online control.
Contributors
Viet-Anh Le, Renukanandan Tumu, and Rahul Mangharam.