Sampling based model predictive control with diffusion score function

Model Predictive Control (MPC) has been widely adopted in robotics for planning and control, but its reliance on constrained optimization, simplifying assumptions, and difficulty in handling discontinuous dynamics limit its use in nonlinear and contact-rich tasks. Reinforcement Learning (RL) addresses some of these limitations through model-free policy learning, yet it demands large amounts of data and computational resources. Recently, Sampling-Based MPC has emerged as a promising alternative, offering flexibility in handling complex dynamics and constraints without requiring differentiability, while also enabling parallel computation. Current approaches, however, typically rely on simple Gaussian sampling distributions, which can lead to suboptimal trajectories, susceptibility to local minima, and high variance in solutions. In this letter, we argue for the importance of controlling or learning the sampling distribution in Sampling-Based MPC to improve the balance between exploration and convergence. We introduce a diffusion-model-inspired sampling strategy that adapts the distribution during planning and demonstrate its effectiveness on a contact-rich robotic task. Our results highlight how structured sampling distributions can enhance the robustness and performance of Sampling-Based MPC, paving the way for more efficient planning and control in complex robotic systems.