86. intelligent design space exploration of hardware-accelerated slam algorithms
Name: Quentin Kevin Gautier
Grad Year: 2019
Alric Althoff, firstname.lastname@example.org
SLAM (Simultaneous Localization And Mapping) and 3D reconstruction algorithms are used in many applications, from 3D scanning of archaeological sites to making autonomous robots and self-driving cars. The SLAM problem has been widely studied and applied over the years to the point of offering dense reconstructions in real-time. Yet, no solution is perfect and new methods are still developed regularly. Each proposed solution is often optimized for a specific environment, and offers a different set of tradeoffs between multiple objectives (accuracy, speed, compute power, etc.). Furthermore, each solution can be tuned in various ways such as changing parameter values, swapping internal algorithms, or accelerating it on specialized hardware. Overall, the space of possible solutions is too immense to be explored. Thus, when designing a new application, we cannot find the most optimized solution to our new problem. We propose to adapt an algorithm, Adaptive Threshold Non-Pareto Elimination (ATNE), to estimate the set of Pareto-optimal solutions for SLAM applications. This algorithm is based on active learning, and was originally developed for hardware design, in which ground truth labels are time-consuming to obtain. Designing a software/hardware SLAM system is a similar situation where labels are expensive to get. We have run preliminary experiments in which we can estimate a set of Pareto-optimal solutions with good accuracy by evaluating 10% to 40% of the space of solutions. We are extending and specializing ATNE to run more accurately on SLAM design spaces with multiple objectives, and more specifically focusing on estimating the space of software/hardware co-design solutions.
Industry Application Area(s)
Software, Analytics | Robotics