Alexander Domahidi

embotech GmbH / ETH Zürich

Thursday, March 02, 2017, 11:00

Room 03-009, Georges-Köhler Allee 103, Freiburg 79110, Germany

With increased computational power, improved algorithms and mature software tools, embedded optimization is often considered a viable technology option for on-board, complex decision making in novel application areas such as autonomous driving, energy management or robotics. While the past decade has seen a tremendous increase in computational efficiency of particular algorithms down to microseconds computation time (“the speed race”), we believe that for a wider success of the technology three major components are still behind their potential: 1) usability through suitable programming interfaces, 2) robustness and 3) (some level of) certification of the algorithms as well as the corresponding auto-generated implementations.

This talk will review some of the efforts by the ETH Zurich startup embotech, which specializes in providing commercial software for embedded optimization, to advance the state-of-the-art in some of these areas. In particular, we will describe the stage-finding algorithm behind the new code generation interface Y2F, which is a human-friendly front-end to FORCES Pro, an established code generator for solving convex multi-stage problems on embedded hardware. Next, we will outline some of the design principles in FORCES NLP, our nonlinear interior point method tailored to solving non-convex optimal control problems. Lastly, we will try to incorporate some of our findings from the market in the past 3 years, and discuss open research questions that we feel are important for facilitating a widespread use of embedded optimization technology for safer, more efficient and certifiable systems.