Adaptive Tube MPC: Beyond a Common Quadratically Stabilizing Feedback Gain

A team of researchers, Anchita Dey and Shubhendu Bhasin, has published a significant advance in Model Predictive Control (MPC) with their paper 'Adaptive Tube MPC: Beyond a Common Quadratically Stabilizing Feedback Gain.' The work tackles a core limitation in controlling complex systems like robots or autonomous vehicles, which operate under real-world uncertainty from sensors, models, and the environment. Traditional 'tube-based' MPC methods handle this by creating a large, fixed safety buffer (the tube) around predicted paths, designed for the absolute worst-case scenario. This leads to overly cautious, inefficient control.

The new 'Adaptive Tube MPC' framework introduces a crucial innovation: online learning. Instead of a static worst-case model, the system continuously learns and refines its understanding of parametric uncertainties as it operates. This allows it to progressively shrink the size of the safety tube and reduce the amount of constraint tightening needed—by up to 40% in simulations. Critically, the method also relaxes a major technical assumption required by prior art, eliminating the need for a single, universally stabilizing controller for all possible uncertainties. This makes the approach applicable to a wider range of challenging control problems while formally guaranteeing recursive feasibility, robust constraint satisfaction, and closed-loop stability.