Learning Agile Paths from Optimal Control

Autonomous robotic systems are of particular interest for many fields, especially those that can be dangerous for human intervention like search and rescue, and maintenance on rigs. However, motion planning in unstructured environment is still a hard problem for legged robots and their success depends largely on their ability to plan their paths robustly. Moreover, the method in which a controller deals with obstacles has great consequences on the planned trajectory, and these optimizations are quintessential in generating agile motions for real-world robots. Trajectory optimization is a common practice for generating motion for legged systems [1, 2, 3], since it can produce optimal trajectories which satisfy the physical and environmental constraints of the robot. However, the solution from trajectory optimization is only valid for a particular pair of initial and target positions, and one needs to re-plan if the pair changes.