Scaling Robot Motion Planning to Multi-core Processors and the Cloud

Imagine a world in which robots safely interoperate with humans, gracefully and efficiently accomplishing everyday tasks. The robot's motions for these tasks, constrained by the design of the robot and task at hand, must avoid collisions with obstacles. Unfortunately, planning a constrained obstacle-free motion for a robot is computationally complex---often resulting in slow computation of inefficient motions. The methods in this dissertation speed up this motion plan computation with new algorithms and data structures that leverage readily available parallel processing, whether that processing power is on the robot or in the cloud, enabling robots to operate safer, more gracefully, and with improved efficiency. The contributions of this dissertation that enable faster motion planning are novel parallel lock-free algorithms, fast and concurrent nearest neighbor searching data structures, cache-aware operation, and split robot-cloud computation. Parallel lock-free algorithms avoid contention over shared data structures, resulting in empirical speedup proportional to the number of CPU cores working on the problem. Fast nearest neighbor data structures speed up searching in SO(3) and SE(3) metric spaces, which are needed for rigid body motion planning. Concurrent nearest neighbor data structures improve searching performance on metric spaces common to robot motion planning problems, while providing asymptotic wait-free concurrent operation. Cache-aware operation avoids long memory access times, allowing the algorithm to exhibit superlinear speedup. Split robot-cloud computation enables robots with low-power CPUs to react to changing environments by having the robot compute reactive paths in real-time from a set of motion plan options generated in a computationally intensive cloud-based algorithm. We demonstrate the scalability and effectiveness of our contributions in solving motion planning problems both in simulation and on physical robots of varying design and complexity. Problems include finding a solution to a complex motion planning problem, pre-computing motion plans that converge towards the optimal, and reactive interaction with dynamic environments. Robots include 2D holonomic robots, 3D rigid-body robots, a self-driving 1/10 scale car, articulated robot arms with and without mobile bases, and a small humanoid robot.Doctor of Philosoph