Improving planning speed and quality

	Topic	Paper page
	Learning transition feasibility and "fuzzy" planning	Learning-assisted multi-step planning
	Adapting motion primitives to new configuration spaces	Using motion primitives in sample-based planning for a humanoid robot
	Optimizing planner behavior	Coming soon.

Improving planner speed

There are several ways to improve planner speed. A learning-based approach is presented above. A lazy planning approach is used in the other legged locomotion papers. My thesis will present a methodical framework that uses learning and probabilistic reasoning to select and order subproblems. This work could potentially be extended to self-tuning software for problem solving.

Improving motion quality

Contact-before-motion approaches typically produce poor quality motions, because they are more concerned with feasibility than quality. Current implementations are quite slow as well. By contrast, motion primitive-based approaches have much higher motion quality and plan faster. A motion primitive is a short, local, high-quality motion designed for a specific situation. For example, for a humanoid robot, a motion primitive might be a motion that takes a single step on flat ground. Then, a search can be performed over all motion primitives given in a library. This type of planner is appropriate when a small library of motion primitives is sufficient for the robot to cover all of the terrain of interest. For example, on wide open flat ground, a forward step, right turn, and left turn should be sufficient.

However, it is impossible to create enough motion primitives to cover all useful motions that a robot might perform, especially in in rough terrain or highly cluttered environments. Furthermore, given a large library of primitives (containing hundreds or thousands of primitives), it is impractical to try every primitive to see if it works, and even more impractical to search using such a high branching factor.

I have done some work generalizing motion primitives to new terrains. This will allow a motion primitive library to remain small, even for highly uneven and irregular terrain. I am also investigating integrating motion primitives with a contact-before-motion framework, and methods for selecting appropriate motion primitives.