Autonomous Vehicles with Quantum Computing - Taking up the Challenge

Recently The Quantum Insider announced the 15 finalists for the Airbus and 宝马 Quantum Computing Challenge 2024. I am very happy to be part of one of the finalist teams and humbled by my amazing team mates Ulrich Wurstbauer , Slawomir Folwarski , Angan M. , Michael Falkenthal , Alejandro Gómez Cadavid and Anton Simen . The team combines the autonomous driving, machine learning (ML) and quantum machine learning (QML) expertise of QuanIT, DXC Technology and Kipu Quantum . As we are participating in the 'Quantum-Enhanced Autonomy' challenge, I want to break down this very interesting quantum use case in this weeks 'Entangled Threads' newsletter.

Getting autonomous aviation and ground vehicles into production is a tricky task. Especially on streets you have to deal with a lot of factors that have to be considered. Other vehicles of different shapes and sizes, pedestrians, animals, construction sides , various other obstacles and even new, so far unknown traffic participants (who would have thought about e-scooters 10 years ago?). This leads to a practically never ending list of potential situations an autonomous vehicle could find itself in and would have to react to. And, it is not enough to get it right most of the time or would you drive in a car that would only be 99% or even 99.9% of the times right? A 1% or even 0.1% of crashing doesn't sound like a bet that I would like to make. The car manufacturers would most definitely not deal with the connected liabilities.

This means that autonomous vehicles have to be trained on a wide range of different scenarios. The best way to do this is of course the training on real situations by using videos, pictures or even exposing the autonomous vehicles directly to these situations. This is easy as long as you are only training on scenarios that happen all the time. But, what is with the one in a billion scenarios? Of course, you want your self-driving car also to know what to do in these situation, but getting real life training data of hundreds or thousands of examples of a similar scenario in various weather and light conditions is basically impossible. That is why most developers of autonomous vehicles rely on synthetic data sets to fill the gaps.

For generating such synthetic data sets you have several options. The most straight forward way is to generate the data fully synthetic using gaming engines. The benefit is that it allows to generate various scenarios quite freely and altering weather and lights conditions can be done using a physics engine. The problem here is the lagging realism and the large amount of computational resources needed. That is why over the last years the use of machine learning to create synthetic data sets was heavily researched. In this approach an algorithm is trained to change weather and light conditions or the exact scenario to generate a much bigger and more divers data set from only a few real data points.

In classical machine learning there are several possible algorithms that can be used, like Variational Autoencoders, Generative Adversarial Networks or Diffusion Models. All of them having different benefits and drawbacks. The hope, backed by recent research, is that quantum machine learning can be used to generate more realistic and diverse data sets from a smaller amount of real examples compared to the classical approaches. This would allow developers of autonomous vehicles to include rarerer and rarerer scenarios into the training of their models, which in return makes autonomous vehicles safer to use.

I hope you excuse the fact that I can't share technical details of how such implementation can be done and what benefits it has, as the challenge is still on-going, but I promise to keep you updated here on the 'Entangled Threads' newsletter about our progress in the challenge and, if you are interested, will also give a glimpse into the 3 other challenges. Let me know in the comments below. Until then, read you next time.

P.S.: Thanks to Ulrich Wurstbauer for the input!