CoCo: Online Mixed-Integer Control via Supervised Learning

Many robotics problems, from robot motion planning to object manipulation, can be modeled as mixed-integer convex programs (MICPs). However, state-of-the-art algorithms are still unable solve MICPs for control problems quickly enough online use and existing heuristics typically only find suboptimal solutions that might degrade performance. In this work, we turn data-driven methods present the Combinatorial Offline, Convex Online (CoCo) algorithm finding high quality MICPs. CoCo consists of a two-stage approach. offline phase, train neural network classifier maps problem parameters (logical strategy), which define discrete arguments relaxed big-M constraints associated with optimal solution problem. Online, is applied select candidate logical strategy given new parameters; applying allows us original MICP optimization We show through numerical experiments how finds near arising in 1 2 orders magnitude speedup compared other approaches solvers.