Building motion models that are both discrete and continuous

Robot motion has continuous properties, such as control effort and distance to the goal, but also discrete properties, e.g. sequence of contacts or an order in which to manipulate objects. Searching for solutions across both domains is often computationally costly but it allows us to define generic problems like planning contact sequences for legged robots, and manipulation of objects on mobile or fixed base robots. Vlad will discuss the formulation of a problem that is both discrete and continuous, and he will present ideas about how to solve this type of problem efficiently by exploiting initialization and heuristics computed using a motion model learnt offline from simulations and demonstration.