Recent technology advancements provide wide accessibility to an unprecedented level of computational power, making it possible to employ numerical methods to attack very large scale problems, such as the trajectory optimization of a humanoid robot taking into account its full dynamic model. In most of the cases, however, these problems are still too hard to be solved at run time and, on the other hand, offline libraries of solutions typically present exponential complexity in the problem size.
This full-day workshop aims to cover recent advances in how offline and online techniques can be effectively fused to develop hybrid algorithms capable to outperform the two extreme methods individually. The discussion will address, on the one hand, the possibility to combine theoretical guarantees (such as optimality and stability) with real-time computations and, on the other hand, the challenges to experimentally employ these techniques.
Convergent contributions from different fields (optimal control, machine learning, convex and nonlinear optimization) will be brought together with special focus on some of the hardest open problems in robotics: contact-aware optimal control and motion generation of largely under-actuated systems. Particular emphasis will be given to the application in the fields of locomotion, manipulation, and control of soft robots.