Motion Planning and Control in high dimensional continuous state-action spaces
Controlling legged robots to allow them to move as naturally and efficiently as animals do is an open research problem. The traditional approach of kinematic planning plus trajectory tracking does not produce plausible motor behaviors and limits the robot ability to perform complicated motions.
Instead, using the dynamics of the robot to generate state and action plans is demonstrating considerable potential. When the dynamics is considered during the planning phase, more fluent and efficient motions emerge, reaching the limits of the robot's physical capabilities.
In my PhD Thesis "Learning rest to rest motor coordination in articulated mobile robots" I worked on the use of machine learning to generate whole body motions in floating-base robots.