Risk-averse and nonlinear MPC: parallel numerical algorithms and applications to autonomous heavy equipment

To this end, we devise SPOCK, a massively parallel algorithm for solving RAOCPs and implement it using CUDA C++ for GPU-enabled hardware. We benchmark the implementation with a Dolan-Moré plot and test a realistic application based on a networked control system with random time delay.

The benchmarking results show that SPOCK outperforms current state-of-the-art interior-point solvers for solving large-scale RAOCPs.<br><br>The other objective of this work, that is, heavy equipment navigation, turns our attention to the case of nonlinear dynamics. We create a hierarchical nonlinear MPC controller for obstacle avoidance, with special emphasis on heavy equipment, particularly skid-steer track loaders. In order to improve the controller solve times, we develop gradgen, a new package for automatic differentiation, that is fast, memory-safe, and has a low memory footprint.