AI in Automotive
The State of Self-Driving Cars for Everybody🚶🚘
With rapid advances in autonomous driving technology, why aren’t we using self-driving cars today?
Making a car drive itself seems trivial nowadays. The hardware, software, and algorithms are arguably sophisticated enough to allow cars to autonomously navigate through streets. For proof, refer to the video below showcasing Waymo’s current autonomous vehicle capabilities.
Volkswagen partners with Apple to develop self-driving buses and showcases their self-driving car SEDRIC. Waymo tests their self-driving fleet in a closed beta version in Phoenix, NV. Tesla boasts the autopilot feature for all models. Drive.ai announced their first publicly available self-driving bus in Frisco, Tx. It’s hard to find a car manufacturer who doesn’t showcase work of self-driving cars.
Autonomous vehicles have the chance to reduce the global accidents rate and save the lives of millions of people. They have the potential to free drivers of the tedious duty of driving. They will improve the security of pedestrians and bicyclists.
If cars can do that and provide so much benefit, why aren’t we reading this article in a self-driving car right now?
Mass-producing self-driving cars (SDCs) faces distinct challenges. Solving these challenges are important to car manufacturers around the world. This post explores and explains six infrastructural challenges to making SDCs available for everyone.
Developing SDCs is divided into six categories, as defined by the standardization institute SAE International. Level 0 defines no autonomy at all, while Level 5 offers absolute autonomy to its passengers. The illustration below explains the capabilities of each automation level. Today, most modern cars are offer functionalities between Level 2 and 3.
To reach Level 5 autonomy and enjoy personal freedom on drives, the following obstacles wait to be overcome.
Let’s drive to the point. 🚗 🚙 🚘 🚖 🚍
1. Data Storage
Self-Driving Cars generate a plethora of data. According to Intel, autonomous vehicles will create 4TB of raw data in a single day. Bring 2,500 vehicles on the road, and you quickly generate a Petabyte of data each day. Also, have you ever tried searching through Petabytes of data? It’s no fun. The sheer scale of the data collected requires elaborate data storage architectures.
Some bits of data also need to be stored in the vehicle. Defining which data is stored is one part. Video is responsible for the vast majority of data generated. Compressing or zipping the camera images is not possible for every use-case.
A car for the mass market achieves about 5% operating margin on a car which costs 20,000$. This yields about 300$ profit per car. Convincing product owners to spend 100$ per car on a large terabyte hard-drive to store data in the vehicle is hard.
2. Data Transportation
Once you’ve generated Terabytes of data, you need to extract it from the car. Most cars don’t enjoy a permanent 5G internet connection yet, so transporting all data over-the-air is challenging.
If you have access to a 50 MBit/s internet connection, you can theoretically upload around 4TB of data in 24 hours. If you have many cars, you need to increase the bandwidth linearly. If the amount of generated data increases, the bandwidth of the internet connection needs to increase as well. SDCs easily test the physical limits of the existing internet infrastructure in many parts of the world.
3. Expense of Sensors
A SDC typically perceives its environment through camera images and the radar technology LiDAR. While cameras are cheap, LiDAR is very expensive. A new LiDAR sensor from Velodyne costs about 75k$, which is already more than we are willing to pay for an entire car. Car manufacturers could consider selling SDCs below their cost of production. They could then get the money back through digital services like in-car entertainment, but it’s a risky bet. Waymo claims to have developed their own LiDAR sensors in-house and subsequently decreased its costs by 90%. This could be one way to use LiDARs in mass-produced vehicles.
There are efforts to steer autonomous vehicles only based on camera images, like presented in this paper. Tesla mainly relies on camera images to steer its cars to avoid the high costs of LiDAR. But as long as LiDAR is necessary for most environment perception functions, it will remain an expensive ‘crutch’ for car manufacturers to develop Level 5 autonomous vehicles.
“In my view, [LiDAR is] a crutch that will drive companies to a local maximum that they will find very hard to get out of. Perhaps I am wrong, and I will look like a fool. But I am quite certain that I am not.” — Elon Musk, CEO Tesla
4. Training Data Acquisition
It’s currently estimated that SDCs will have to drive about 100 million miles to have collected enough data to safely navigate autonomously. Waymo finished the completion of its 5th million miles driven in February 2018. Although the pace increases steadily, given the current rate, it’ll take Waymo a long time to drive the remaining miles. Companies also drive millions of virtual miles on a daily basis. But collecting results in the real world is time-intense.
Pay attention to the pace of acquiring data for Waymo. While it took the company 6 years to drive the first million miles, it drove the latest million miles in 3 months. That’s impressive, but there is still a long road ahead.
5. Acquisition of Corner Case Data
After the tragic accidents involving autonomous vehicles from Tesla and Uber, the need for having corner case training data becomes ever more important. Corner cases are situations which rarely happen, e.g. a pedestrian unexpectedly stepping on the street or placing a concrete block in the middle of the street. It’s good to have data from driving on the highway, but much value lies in corner case training data. Waymo apparently tests its autonomous vehicles by randomly jumping in front of it and checking if it stops or not. Current areas of research involve creating corner case data from simulated environments or creating artificial data through text-to-video GANs.
6. Verifying Deep Neural Networks
Understanding why or why not an SDC identifies another car is paramount to convince regulators that these cars are safe enough for public use. SDC’s rely on Deep Learning algorithms, which are notorious for not explaining why they decide one way or another.
In a famous project, researchers trained a neural network to distinguish between wolfs and dogs. The model achieved an impressive accuracy. However, the researchers eventually found out that the neural network learned to detect snow on images since most training images of wolves contained snow in the background. That’s not the conclusion that the network should to draw.
Before regulators allow cars to autonomously drive on public roads, they need to verify that the vehicles are safe.
This list covers infrastructural problems of producing self-driving cars for the masses. This list may not be extensive, but it provides an overview. You’ve seen the problems for self-driving cars for everyone. What solutions do you suggest?
As you see, there is still much work to do to verify the algorithms and expand the data training and storing capabilities of SDCs. If you think that the workaround SDCs is important and interesting, feel free to check out the job postings by the Electronics Research Lab in the San Francisco Bay Area, USA, the Autonomous Intelligent Driving GmbH in Munich, or Carmeq in Berlin.
