Efficient implementation of Randomized MPC for miniature race cars

This paper describes a data-driven Randomized Model Predictive Control (MPC) approach toward autonomous racing of miniature cars. The main challenge in autonomous racing is to drive as fast as possible, without actually exiting the track. Designing such controllers is a challenging task, since (i) the models of the cars are not entirely accurate, and (ii) the time to compute the control inputs is limited to tens of milliseconds. The current practice is to resort to standard MPC, where the safety-related constraints are tightened manually. In this paper, we propose an alternative method based on Stochastic MPC, which naturally introduces robustness by design. We approximate the Stochastic MPC problem by means of randomization, implemented as a sparse Randomized MPC problem. We show that the resulting quadratic program can be solved in the range of milliseconds. Moreover, the sparse Randomized MPC formulation can be interpreted as a standard MPC problem whose constraints are, based on collected data, automatically tightened. We show experimentally that the proposed controller can outperform currently implemented controllers. Our results suggest that Randomized MPC is an attractive alternative to standard controllers, due to its ability to introduce robustness without sacrificing performance.