Abstract
Technical Project
The premise of my capstone research is that there has to be a better way to adjust the level of performance you're demanding from the Cavalier Autonomous Racing Team's autonomous IndyCar during testing without redoing the heavy math. Our raceline, which we strive to follow in a time trial setting without other cars, represents the precomputed optimal route around the track, including where to brake, where to turn and how hard, and what speeds to accelerate to. This is a difficult optimization problem that, with our current software, can take many minutes to compute the solution.
However, when testing on track last summer at Laguna Seca, we could not run at full speed/tire demand as our car is not yet perfect. Our controller is not perfect, our estimate of the tire capabilities may not be perfect, and so inherently we want to be conservative in our driving.
So, I created a system that runs in linear time (relative to the length of the track, essentially computationally free) which can revise the demanded car speed to put lower demands on the car. A key insight is that the difficult math has already computed, based on its own parameters, what it thinks the tires can do, what the engine can do, and what the brakes can do at that particular point on the track, and we don't need to recompute that difficult part in real time.
The algorithm itself is multi-stage, but all in linear time. We look at the optimal trajectory and find the demands being placed on the car at that point. We scale down the demands we're willing to make on the car according to our parameters, which are acceleration utilization, lateral utilization, and brake utilization, such that if our lateral utilization is 70%, for example, we will only put 70% as much force on the tires as we would in total. Then we run a pass that reduces speed to enforce the lateral acceleration constraint. We run a forward pass to ensure the vehicle does not accelerate faster than what the powertrain limit allows, and a backward pass to ensure the vehicle does not have to brake harder than it is capable of.
This technique is also extensible to future useful work. For example, if we can measure tire temperature in real time, an algorithm could revise upward the force the car is willing to apply to its tires as they get hotter, and all this recalculation can happen in real time.
STS Paper
My STS paper looks at a phenomenon I've seen unfold growing up in my hometown of Silicon Valley and recently when Cruise and Waymo were approved to begin offering paid passenger service. San Francisco was the testbed for passenger autonomous vehicles and drew a lot of intrigue and emotion when emboldened AVs clogged up the city streets. A main source of friction was that the approval came from the state of California, and not the city where the testing occurred. Local stakeholders mounted significant backlash, including first responders and community leaders.
In my paper, I argue that public perception of AVs is skewed by an asymmetry in which the follies of human drivers, even though more frequent and more dangerous, are not talked about because they are normal and consistent, and that autonomous vehicles were held in the public eye to a high and, at first, unreachable standard. I also discuss the impact of the redistribution of liability and responsibility from millions of human drivers to large companies that operate autonomous vehicle fleets. This carries both pros and cons, as errors in autonomous algorithms can propagate to potentially thousands of flawed vehicles. But this is also a benefit, as improvements can be made to autonomous drivers at scale in a way that is not true of humans.
I introduce the idea that what happened in San Francisco showed that autonomous vehicle companies are more accountable for their driving behavior than humans often are. Humans are indemnified by the insurance they are obligated to purchase. I reflect on the overall outcome of the autonomous vehicle rollout, the impact it has made and the potential it carries, and how it measures up against what the autonomous vehicle rollout could have been.
Linking the Two
These two topics are linked by my interest in autonomous vehicles. Working on the technical project, I learned how difficult many technical challenges in the autonomy space are and how important physics and math are to efficient, safe driving. I gained more respect for the engineers behind the autonomous systems I describe in my STS paper and learned about both how industry is progressing and where challenges still lie.
