UC Santa Barbara

Simultaneous Localization and Mapping (SLAM) describes a class of problems facing a large and growing field of autonomous systems -- from self-driving cars, to interplanetary rovers, to home automation products. Unfortunately this is a complex task where sophisticated algorithms and data structures are required to navigate a wide range of uncharted environments. Furthermore, most mobile robots need to run these tasks near real-time onboard an embedded controller with limited power and compute resources. To address this problem we explore the stochastic gradient descent (SGD) variant of graph solvers for SLAM and observe a tradeoff between various execution architectures and overall execution speed. Based on these observations, we propose a custom multiprocessor design that relaxes memory-coherency constraints between parallel cores while avoiding divergent behavior. We introduce a specialized streaming-tree interconnect that provides increased performance while using fewer resources compared to state-of-art GPU/CPU implementations of SGD. Finally, we discuss applications of unconventional architectural paradigms like over-provisioned “dark processors” and specialized data partitioning that provided a unique performance advantage for our particular design.