| dc.description.abstract | Planning motion is an essential component for any autonomous robotic system.
An intelligent agent must be able to efficiently plan collision-free paths in order to move through its world. Despite its importance, this problem is PSPACE-Hard which means that even planning motions for simple robots is computationally difficult.
State-of-the-art approaches trade completeness (always able to provide a solution if one exists or report none exists) for probabilistic completeness (probabilistically guaranteed to find a solution if one exists but cannot report if none exists) and improved efficiency. These methods use sampling-based techniques to design a sequence of motions for the robot. However, as these methods are random in nature, the probability of their success is directly related to the expansiveness, or openness, of the underlying planning space. In other words, narrow passages, complex systems, and various constraints make planning with these methods difficult. On the other hand, humans can often determine approximate solutions for these difficult solutions quickly.
In this research, we explore user-guided planning in which a human operator works together with a sampling-based motion planner. By having a human-in-the-loop, a human can steer a sampling-based planner towards a solution. This strategy can provide benefits to many applications such as computer-aided design and virtual prototyping, to name a few.
We begin by classifying and creating simple models of common user-guided and heuristic-guided motion planning methods. Our models encompass three forms of user input: configuration-based, path-based, and region-based input. We compare and contrast these approaches and motivate our choice of a region-based collaborative framework. Through this analysis, we gain insight into user-guided planning and further motivate methods that harness low interface complexity and work entirely in workspace, which is most natural to a human operator. Further, we extend the theory of expansiveness to analyze the various types of user inputs.
Our novel region-based collaboration framework takes advantage of human intuition by allowing a user to define regions in the workspace to bias and/or constrain the search space of a sampling-based motion planner. This approach allows a user to bias a high dimensional search with low dimensional input, supports intermittent user hints, and empowers a user to customize motion solutions. Finally, we extend region steering to both non-holonomic robotic systems and a human-inspired approach to motion planning.
Our results show that this region-based framework can aid many variants of sampling-based planning, reduce computation time, support solution customization, and can be used to develop advanced heuristic methods for solving motion planning problems. We provide experiments exemplifying our approach in planning motions for complex robotic applications such as mobile manipulators, car-like, and free-flying robots. | en |
