UC San Diego

Deeply buried within the jungle of Central America lie the remains of the ancient Maya civilization. In order to reach these old ruins, archaeologists dig tunneling excavations spanning tens of meters underground. Unfortunately, most of these excavations must be closed for conservation purposes, and it is therefore crucial to document these findings as precisely as possible. Many solutions exist to create a 3D scan of these environments, but most of them are too costly, difficult to deploy, or do not provide real-time feedback.

A possible solution to create a 3D mapping system overcoming these problems is to use Simultaneous Localization And Mapping (SLAM) algorithms, combined with low-cost sensors and low-power hardware. However, the combined complexity of software design and hardware design represents an immense challenge to implement a system optimized for all requirements. The vast pool of possible designs and the multiple, often conflicting objectives contribute to produce design spaces too complex to be explored manually.

In this thesis, we explore the complexity of designing SLAM applications using various types of hardware. First, we manually evaluate SLAM algorithms for 3D mapping, and specifically optimize one SLAM algorithm on an FPGA hardware. Then we expand our exploration to a larger space of designs, and generalize this problem to the design of all complex applications that require a lengthy evaluation time.

We develop several learning-based methods to help designers finding the best combinations of optimizations that maximize multiple objectives. By using a smart sampling algorithm, and removing the need of selecting a specific regression model, we can help users largely decrease the number of designs to evaluate before reaching an optimal architecture.