Driverless Tech Finding New Life in Old Data

By Graham Jarvis

---

20th August 2020

Testing connected and autonomous vehicles (CAVs) in the real world is said to be the gold standard for being able to collate data to improve them.

However, for much of 2020 the lockdown has led to the use of simulation modelling to ascertain how to improve them and, to a degree, this often means that old data has been needed. It’s important because it can be used to track when and why problems have occurred, enabling changes to be made, new algorithms and new scenarios to be explored with some understanding of how to improve them in hindsight.

Virtual training

Hayden Field writes in his article for MIT Technology Review, Self-driving cars are being trained in virtual worlds while the real one is in chaos: “For safety reasons, autonomous vehicles typically have two operators apiece. That’s a no-go in the age of social distancing, and leaders of autonomous-vehicle companies knew they’d have to mothball their fleets.” For example, he reveals that “General Motors-owned Cruise has relegated 200 vehicles in San Francisco and Phoenix largely to the garage.”

“This has happened to comply with the regulations that were imposed to stop the spread of Covid-19, despite real-world testing being the gold standard for collecting data and for collecting data and improving the cars’ ability to operate safely.  As fully autonomous vehicles are largely experimental at the moment, and with no way of getting on the road soon, the self-driving operations faced the risk of becoming cash-intensive “with no path toward fielding a product anytime soon.”

Filling the void

So, to fill the void and to make the best of a bad situation, Field says several companies traded road tests for “delving deep into their algorithms and simulator” to find new uses for the hours of data they’d already collated. Their focus turned to data labeling, 3D mapping and identifying scenarios that had been previously overlooked from previous road tests.

All of this data can be used, even though if it is historical data, to train their systems. Vehicle operators have even been helped to “transition into data labeling, equipping them with new skills that will likely come in handy when they resume their former roles,” he adds. New tools have also been built to enable the annotation of driving data. This is creating a new life for old data. Combined with new data it can be used to test scenarios, software and hardware to improve vehicle performance and safety.

For example, autonomous trucking firm Embark Trucks annotated four years of data. The software they used showed truck drivers several images of different scenarios, which they were asked to determine whether or not they were noteworthy. They were also asked how they would handle each situation, based on their own experience.

Mining data

A similar approach has also been taken by Aurora Innovation. Its vehicle operators are collaborating with its triage and labeling teams. In an email to MIT Technology Review, the company’s CEO and co-founder Chris Urmson, wrote that this approach is enabling them “to mine our massive collection of manual and autonomous driving data for additional interesting on-road events that can be turned into virtual tests”. Operators are also now more aware of how the data they collate is used offline. This data will enhance the ability of the operators to do their jobs when time comes to go back on the road.

Autonomous vehicle companies have had to become creative to overcome the obstacles of not being able to test their products in real-life. Embark, for example, put its money into software for testing hardware components offline – an investment which Brandon Moak, CTO and co-founder of Embark Trucks, feels will put his firm in good stead. The software his team developed permits hardware components to be tested offline to assess the vehicle control systems, namely “responsible for sending physical commands, like how fast to turn the steering wheel,” explains Field.

Lyft: back on the road

While the situation remains uncertain, particularly with regards to Covid-19 infection rates being extremely high in the United States, Lyft has begun to once again test its vehicles on public roads in California. Robert Morgan, Director of Engineering and Sameer Qureshi, director of product management at Lyft Level 5 – the autonomous vehicle division of Lyft, write in their article on Medium, Continued Momentum Through Simulation: “We’re excited to announce that our autonomous vehicles (AVs) are back on the road and that, during the shelter in place, we continued to make progress by doubling down on simulation. Simulation is an important part of our testing program, enabling us to test beyond road miles.”

The nevertheless add: “If reliant upon on-road miles, it may take some number of billions of miles to test everything. Simply put, the scale makes it impractical to rely only on road miles. Therefore, we supplement our on-road testing with simulation, which gives us a cost-effective way to create additional control, repeatability and safety. It also allows us to test our work without vehicles, without leaving our desks, and for the last few months, without leaving our homes.”

Once the lockdowns are over, more and more autonomous vehicle testing on track and on public roads is likely to occur. Autonomous vehicle developers need to test the vehicles in real-life conditions and apply the lessons from both to improve the development of CAVs. To do this they can use the data they collated during lockdown to test it in real-life.

UK trials unveiled

To this end, in the UK, plans were unveiled for a 180-mile test road for autonomous vehicle trials in the Birmingham, West Midlands, area. The road, it is hoped, will be available to car manufacturers and technology firms later this year, 2020 – once Covid-19 has been brought under control.

This will need to occur to a reasonable level because the first vehicles used on the route between Birmingham and Coventry won’t be driving themselves in the early stages of the road trials. They will have a human driver, and occasionally – and this is where social distancing could be an obstacle, they will need a second person accompanying them to monitor their performance.  So, there are still some potentially challenging times ahead, and as a result they may still need to find new life in old data.