A ‘rage’ to learn: Red Bull Advanced Technologies uses OCI to test AI for future racing stars

To make the leap to the pinnacle of motorsport, up-and-coming racecar drivers need much more than a typical athlete’s intensity and focus.

“We look for what I would call the ‘rage to master,’ which is this burning desire to learn and to be better the next time they drive around the track,” says Guillaume Rocquelin, who coaches and develops young drivers as the head of Red Bull’s driver academy.

“It's being very, very hungry for any kind of data and analysis we can provide them,” Rocquelin says. “That's the key attitude we're looking for. Now that doesn’t always translate into physical ability, but that's the starting point.”

No one knows more about making that leap to the highest level of racing than Rocquelin, who was the race engineer with four-time Formula 1 world champion Sebastian Vettel before becoming head of race engineering and then moving to his current role. The Red Bull Driver Academy is one of the sport’s best training grounds—7 of the 20 drivers on today’s F1 grid are graduates of the program.

Rocquelin has seen areas for development in the teaching tools that coaches like him have. While great young drivers must be voracious learners, a coach has very limited ways to show exactly why one driver goes faster than another. Precisely when did a driver brake, accelerate, or down shift? What angle did they take into and through a turn? A driver can only watch and try to replicate what they see in someone going faster.

Red Bull Advanced Technologies—which applies high-performance engineering to motor racing and looks for uses of that tech in other industries working—is working on this challenge of creating better young driver training tools. Engineers at Red Bull Advanced Technologies are working with Oracle data science experts to explore how machine learning, cloud computing, and data visualization could work together to create a more valuable training experience for these data-hungry athletes.

“In any type of coaching environment, the tools are just the start of the conversation, and a higher quality of tool will mean a higher level of conversation,” Rocquelin says.

In any type of coaching environment, the tools are just the start of the conversation, and a higher quality of tool will mean a higher level of conversation.

Guillaume Rocquelin Head of Red Bull Driver Academy

Oracle’s data science team is using Oracle Cloud Infrastructure (OCI) to refine the algorithms used in self-driving cars—known as simultaneous localization and mapping, or SLAM, algorithms—and apply them to analyzing racing video, initially from the team’s esports drivers. Once successful, the team expects to be able to input video of a driver’s session into an application, run machine learning analysis of that footage, and glean new insights into how to improve their lap times.

The tool is still in the early development stage. Oracle’s data scientists encountered a challenge, for example, when they applied the self-driving car algorithms to race cars instead of a standard sedan. “The laws of physics are very different when it comes to racing,” says Jigar Mody, vice president of artificial intelligence services at Oracle. Here’s how OCI is helping the teams address an unsolved AI challenge.

To get started, Red Bull Advanced Technologies gave the Oracle data science team video from its esports simulators for analysis, and the Oracle team applied the SLAM algorithms to assess where a car is on the track. They hoped the output would provide the basis of data needed for analysis.

The problem: When the team first applied SLAM to race video, the predicted location was off by a half kilometer. These algorithms weren’t built to understand a vehicle that moves at typical top speeds of 320 km/h (200 mph) and can stay on the road while pulling 5Gs of lateral acceleration in a turn. An accurate AI model is needed for the data analysis systems they envisioned, so the Oracle data scientists got to work refining the model.

Such precision is important. “They’re very accurate in how they drive, so 20 centimeters positional accuracy is very necessary for algorithms to be at all useful,” says Dr. Alberto Polleri, Oracle Chief Data Scientist and the AI expert guiding the project. “And for the angles of very few degrees that describe car direction, it must be accurate to within less than a single degree."

The Oracle data science team heavily uses OCI graphics processor units (GPUs) as on-demand compute power to support the large computing loads used in AI modeling and testing. The team ingests video into OCI, then uses convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to process the images; then they use OCI to test various parameters and fit the AI model, seeing how well the results match on-track reality.

This process of fitting the model is the most computationally intensive part, so having computing capacity as an on-demand, variable cloud resource on OCI is essential. Some models might take days to run. Sometimes the team will start a test and see it’s not improving the model and shut it down, or they might run different models in parallel. “We do hundreds of experiments a month that take multiple days each,” Polleri explains.

Here’s how data flows through the architecture on OCI:

It starts with ingesting a video footage into OCI.

Then data flows to three parallel pipelines, or workflows, to assess the visual odometry (the car’s velocity and orientation); on-track location; and car controls (the steering wheel and wheels). These three workflows use extensive OCI Compute, including GPUs.

Once the AI models are refined, Red Bull Advanced Technologies hopes to have a tool running on OCI that lets it input video and get deep analysis on what a driver did differently from lap to lap, or what one driver does differently than another.

This kind of algorithm research also could prove valuable in applications beyond racing, to areas such as robotics and autonomous vehicles—in any application where it would be useful to predict where the next movement of an object is. While most use cases won’t involve the 200-mph pace of a race car, these refinements made in racing should also help at more common velocities. “The technology that works well at high speed does phenomenal at lower speeds,” Oracle’s Mody says.