Plain English Explanation

Problem: Let’s say you have a cyber-physical system, like a drone that is surveying a certain area of land. The drone has both a "cyber" resource (a camera that takes images as a given rate) and a "physical" resource (an engine that flies at a certain speed). The mission objective is to survey the most land possible. For the drone to survey the most area, you want the engine to be running as fuel-efficiently as possible (usually somewhere in the middle of possible speeds), and to take as few pictures as possible while still photographing everything (the lowest rate that will not miss any land). These two resources are interrelated: if the drone flies faster, it needs to take pictures more frequently, and if it goes slower, it needs to take pictures less frequently, but may waste cyber capacity. Traditionally, optimizing a system like this involves constructing a single “cost function” that captures the relationship between the resources. Given an engine speed and an image rate, the cost function will output a number representing the overall system performance in achiving the mission objective. Optimization techniques can be used to follow the smooth curve of the cost function to its peak (or valley), which tells you the combination of engine speed and image rate needed to achieve the maximum system performance (or minimum system waste). However, such optimization is difficult in the case of cyber-physical systems because it is hard to create a smooth cost function out of both discrete (non-smooth) and continuous (smooth) components.

Problem Impact: Failure to account for the relationships between cyber and physical resources means having less efficient missions. Less efficient missions can mean different things for different applications. For the surveying application above, that could mean unnecessarily wasting fuel and having to return to refuel more frequently than necessary (costing money and time), or missing areas in the survey photos, which could cost money and time to redo later, or potentially cause other issues.

Solution: We made a multi-disciplinary cost function and used a numerical methods approach to optimization that does not need the cost function to be smooth.

Evaluation: We created a toy surveying application (the above image surveillance task) and implemented it using a TableSat (table-top satellite). We used the TableSat testbed to run experiments with different parameters.

My Contribution: I did the majority of the TableSat programming and evaluation and contributed text to the paper.

Reflections

This was my first-ever paper, written my first semester as a PhD student. The first author was a very senior PhD student who went on to become an aerospace professor. I’m thankful for his mentorship. It’s wild thinking back on how little I knew what I was doing then, and how much more I know now. I has just started in embedded systems in my previous job. I really enjoyed hacking on the TableSat. At the time my focus area was theory, so the aerospace lab felt like home.

The evaluation in this paper wasn’t so great due to open-loop control on the TableSat’s fan speed. The fan speed was set using PWM. We made the assumption that if you set the fan speed, the TableSat would reach a consistent equilibrium rotational speed. This turned out not to be the case, because the TableSat was balanced on a non-stainless steel screw. As the TableSat spun, it dug a pit into the tip of the screw and started experiencing variable friction as it wobbled. This caused inconsistencies in the TableSat’s rotation speed that made our results noisy. I still have that screw to this day, as a reminder that if the success of your entire project rests on a single point, you better make sure that point is hardened.