In the past decade, the development of dynamic legged robots has gained significant traction and has produced remarkable results. Modern bipeds and quadrupeds are able to traverse difficult terrain while using various gaits and movements like walking, trotting, or jumping. Controlling those robotic systems can be a challenging task that often requires a lot of computational resources. Many modern control architectures employ state-of-the-art strategies like Model Predictive Control. These controllers allow for complex behaviors and movement patterns, while also accounting for dynamic balancing. This is made possible by mathematical optimization, or in many cases, by quadratic programming. These quadratic optimization problems can be solved by numerical methods, implemented by QP-Solvers.
This research focuses on comparing different QP-Solvers on different computer architectures, namely x86 and ARM. The aim is to evaluate the efficiency of each solver based on their respective solve speed and CPU power consumption. Combining these measurements into a Solve Frequency per Watt metric makes it possible to compare QP-Solvers between different computer architectures easily. The specifications of the onboard computer of a dynamic legged robot can have a great influence on its performance. Having knowledge about the efficiency of the used QP-Solver can therefore help in enabling tasks that are more complex or require longer endurance.