Embedded Learning-based Model Predictive Control for Mobile Robots using Gaussian Process Regression

This paper proposes a learning-based Model Predictive Control (MPC) strategy for mobile robots. The strategy relies on Gaussian Process Regression (GPR) to improve the robot's model based on measurement data, therefore avoiding the need for an accurate model. More specifically, the kinematics are assumed to be known and GPR is employed to capture the unknown robot dynamics. Within a cascaded control structure, the resulting full system model is used to generate velocity commands by means of MPC. This allows the dynamics to be taken into account during velocity reference generation, while keeping a fast inner-loop controller to track the velocity reference. The variance provided by the Gaussian process is used to tighten constraints for robust constraint satisfaction. Incremental Sparse Spectrum Gaussian Process Regression (ISSGPR) and Real-Time Iteration (RTI) are employed to facilitate real-time online learning and ensure implementability on embedded hardware. Simulations and experiments are performed to show the effectiveness of the proposed approach. Although we focus on the application to Wheeled Mobile Robots (WMR), the ideas provided here are applicable in other contexts where the kinematics are known and there are unmodelled dynamics.