Phase-indexed ILC for control of underactuated walking robots

We propose a method for learning the nominal control input for virtual-constraints-based walking robots. The key problem in applying iterative learning control (ILC) to these systems is that the iterations are only approximately periodic. We develop a modified form of ILC that indexes previous iterations by a phase variable (a function of the state variables) rather than time. We show in experiments that the proposed method outperforms time-indexed ILC and a hybrid “resetting” form of ILC in terms of tracking error reduction and stability.