Model Predictive Control (MPC): A Breakdown of Recent Research
A team of researchers from Carnegie Mellon University recently released a comprehensive analysis of the convergence characteristics of Model Predictive Path Integral Control (MPPI), a popular sampling-based MPC technique.
MPPI’s Main Characteristics
The study aims to understand MPPI’s convergence behavior in cases where the optimization is quadratic, such as time-varying linear quadratic regulator (LQR) systems. The study found that MPPI shows at least linear convergence rates under certain circumstances and expands to include non-linear systems as well.
Introducing CoVariance-Optimal MPC (CoVO-MPC)
Based on the research, the team has developed a new sampling-based maximum probability correction method called CoVariance-Optimal MPC (CoVO-MPC) which optimally schedules the sampling covariance to maximize the convergence rate. This method has been empirically tested and has been shown to outperform regular MPPI by 43-54% in both simulated environments and real quadrotor control tasks.
The Study’s Primary Contributions
The study provides insights into MPPI’s convergence behavior and introduces the CoVariance-Optimal MPC (CoVO-MPC) algorithm, showcasing its notable gains in real-world applications.
In conclusion, this study advances the theoretical knowledge of sampling-based MPC and presents a unique technique that shows significant improvements in real-world applications.
For more information, check out the paper and Github. And don’t forget to follow the researchers on Twitter for updates and join their various social media groups and newsletter.