Self-Driving Car

Final Results

Self-Driving Car

Advisor: Sherri Goings

I. Background

Just a few weeks ago a self-driving car travelled over 3,000 miles across the country from San Fran to NYC, almost completely without any human intervention. In fact the car was in full automation mode 99% of the trip, with humans taking over the controls only when exiting the highway to drive on city roads. Clearly there are many issues that will need to be addressed before self-driving vehicles become the norm on our roads, but the technology that makes it possible for a computer to safely drive a car without any human intervention isn't just something in sci-fi stories anymore, it already exists.

The recent cross-country driving car was equipped by an English tech company named Delphi, who designed a set of sensors and auto-driving software that could be added to any car. The setup included 10 radar instruments, 3 cameras, and six laser range detectors, as well as of course lots of software to interpret all of the incoming data and determine the appropriate action for the car to take. This car had to drive itself using only local data from its own sensors, though this still meant its "brain" had far more information than a human driver ever gets, which is one of the big potential benefits of autonomous vehicles. Imagine how much safer human drivers could be if they could "see" in every direction at once!

Even more promising is the future possibility of networked self-driving cars that can share data in order to "see" what's happening around blind corners or miles ahead. The hardware already exists to share this type of data among self-driving cars, but the more difficult aspect is creating the software to best use the abundance of information. Several companies and researchers are competing (and/or cooperating) to invent and improve methods for self-driving cars, and we can expect to see major development in this area over the next few decades.

II. The Project

The goal of this project will be to design and analyze algorithms to control an autonomous vehicle. This vehicle and data will exist in a software simulation only, as it would be decidedly impractical to create an actual physical self-driving car. The focus of this project will be on the "brain" of the car itself, not the development of a physics simulator or impressive graphics. The steps of the project will look something like the following:

Choose an out-of-the-box physics simulator, implement a very simple car with semi-realistic basic controls, such as accelerating, breaking, and turning.
Add basic simulated sensors such as laser distance and line detection to keep the simulated driving car in its lane.
Add sensors and capabilities to the car so that it can navigate a more complicated environment, though still assuming that the car is the only moving object in its world. This should include at least camera sensors to handle basic traffic signals such as stop signs and stoplights.
Add other well-behaved traffic that your car must take into account, at least to avoid collisions, ideally to navigate around/through.
If you complete all of the above, there are lots of interesting directions to explore! For example:
- How can you use external data from a network of cars to improve the performance of your self-driving car?
- How well can you make your car react to sudden unexpected events, such as another car running a red light, or braking suddenly, or the myriad of other things an actual driver encounters on the road?
- Can you handle more realistic sensor data with noise?

III. References

delphi.com/delphi-drive
http://www.templetons.com/brad/robocars/simulator.html (comprehensive list of open-source simulators that support driving a car)

IV. Recommended Experience

Algorithms would be very useful, and any experience in physics or working with simulations would help as well.