Blog post

Challenges You Can't Escape While Training Algorithms for Autonomous Cars

Intellias Delivery Director for Automotive Oleksandr Odukha speaks on data analysis and automotive artificial intelligence use cases

November 29, 2019

4 mins read

The automotive industry continues to face new challenges. For those who are ready to innovate and shift towards greater changes, it represents an exciting time with new opportunities. We had a chance to talk with Oleksandr Odukha — Delivery Director at Intellias. Responsible for automotive portfolio projects and ML domain, Oleksandr shares insights on why data is “the new oil” and what challenges you can’t escape while training algorithms for autonomous cars.

What are the key trends driving the growth in AI and Machine Learning?
For starters, I would say that data is at the core of nearly every business decision made. We witness that today, more industries consider data as an integral part of their business high-value performance. This especially applies to the automotive sector. For example, the key differentiator for the automotive industry will soon switch from “what kind of software companies use” to “what data do companies have”.

What are the current ML and AI advances in the automotive domain?
Automotive artificial intelligence is all about data. But let’s be honest: some companies don’t possess enough data, some companies have too much data. Finally, some companies still struggle to understand how to apply the existing data to benefit their business and products properly. To top it off, all of these companies lack the experience and understanding of how to make data work for their business needs and build data-driven strategies. You see, that’s the problem.

There are situations that you can’t simply simulate in the real world. In this case, you have to create simulated environments and use them to build data models.

Today we don’t have enough data, so how to build those data models?
The term “data modeling” carries a lot of meanings. For our purposes, let’s take an example of reinforcement learning on how to “train your car” not to hit pedestrians. To create a testing environment, you will need to build a simulation to fully understand the accidents. This will help to discover the deadliest dangers and test solutions to make a safer car.

How problematic is it to build data models in the scope of autonomous driving?
Autonomous driving projects face significant data challenges. Each autonomous vehicle generates data arrays. For example, take a look at Tesla’s use of data. It’s where the company shines. They actively collect data from thousands of vehicles to predict how the car might perform. Their vehicles are engineered to be the safest cars in the world and they managed to achieve the lowest percentage of car accidents. But if the car injury happens, it’s then a subject of a huge media hype.

In other words, the models used in autonomous driving should be precise, predictable, and fast enough. Thus, to achieve it, you have to do your best to get more of the data and skills you have.

What does the world start to look like when more and more vehicles have all these sensors and are connected?
We can experience a wide range of features associated with the connected car. Many drivers already link their smartphones to their cars. Honestly, I think that in the nearest future there will be fewer cars on the road. First of all, the number of connected vehicles will grow. Connectivity will be the key to using car data to improve safety. Think about airplanes: we know where each airplane on this planet is right now in real-time. I think that at some point it will happen to the automotive sector as well. The cars will soon be able to exchange the data they have, leading to fewer car accidents and a safer world to live in.

How is Intellias contributing to the autonomous driving?
First of all, we have internal research projects that we use as accelerators to speed up development for our clients. Additionally, we work as consultants and focus on real-world issues and ways to solve them. We cover everything from business needs alignment, actual data availability, data preparation model, model evaluation to implementing it in the real world.

As of now, we’ve just finished accelerator for pedestrian recognition and for recasting the environment around the car. Basically, we’ve analyzed the probability of vehicle-pedestrian collisions. The data received allows us to warn the driver to slow down or use an emergency brake.

The automotive sector isn’t just being driven by people. Today, data gets the key role. Autonomous vehicle manufacturers strive to deliver a better customer experience by moving towards autonomous, self-driving vehicles. But as of now, their key focus is on the safety and reliability of cars. The example of Intellias shows how data can help to gather insights that can be used to release better and safer cars on the road. And it’s only the beginning of a new era for the automotive sector and safer roads.

Your subscription is confirmed.
Thank you for being with us.

No votes Thank you for your vote. 20898 c6d0ace301

Thank you for your message.
We will get back to you shortly.