Home > Analysis > Training, not tech, is slowing AV development

Training, not tech, is slowing AV development

Algorithms are the biggest secret behind the autonomous car – but for the foreseeable future, those algorithms need to learn what humans already know. Randi Barshack of Figure Eight talks to Michael Nash about training machines to learn

Algorithms are capable of solving the most complex mathematical equations in nanoseconds. They are part of our everyday lives, crucial to the Internet, Netflix and even home energy management systems.

They will also play a fundamental role in the autonomous vehicle. However, these algorithms need to be highly intelligent to ensure that the vehicles can operate safely and efficiently.

Based in San Francisco, Figure Eight (formerly CrowdFlower) provides a human-in-the-loop (HITL) platform that helps data scientists to collect, clean and label data. This can then be used for machine learning.

Back in June 2017, the company announced that it had received US$20m in funding, most of which was provided by Industry Ventures. The investment has been used to extend the functionality of its platform and hire new data scientists, machine learning experts and engineers.

We’re not really waiting for any additional technology. The bottleneck at this point is the training. It’s just a matter of how much manpower can we throw at this

Speaking to Automotive Megatrends, Randi Barshack, Vice President of Marketing at Figure Eight, provided some more background information on the platform and how it could be vital for the development of autonomous vehicles.

Secret training

“The connected and autonomous vehicle is a huge topic of discussion in both the artificial intelligence (AI) industry and the automotive industry,” she observed. “Some would argue that the newest models on the market are not even cars, but moving computers. The dirty little secret is that these vehicles need to be trained so that the algorithms they use can perform in the real world.”

The misconception here, continued Barshack, is that algorithms are written by engineers, and then simply used in autonomous vehicles to bring an unprecedented level of intelligence. “But an algorithm on its own is just theoretical potential,” she emphasised. “It can avoid hitting a tree or a pedestrian, but it needs to be taught what a pedestrian or a tree is. This seems so counterintuitive to us as humans because its second nature for us, but for a computer, it’s extremely challenging.”

Using Figure Eight’s platform, a person takes an image or a video clip obtained by the various cameras and sensors on a vehicle and then draws boxes around various sections to ensure the algorithm can distinguish between a pedestrian and a tree, for example. The input is then fed back to the algorithm as ‘training data’.

“So we are literally training the algorithm to be more intelligent,” Barshack continued, adding, “The way the algorithm absorbs information is very similar to having a three-year-old child learning from the images that we provide.”

The range of scenarios

However, this education programme can take a significant amount of time and effort, as the algorithms need to be highly accurate to ensure that autonomous vehicles are safe and robust. “If we expect cars to be autonomous then they must be intelligent,” she said. “They must be able to firstly identify a road sign, and then secondly be able to tell what that road sign says so as to react accordingly.”

An algorithm on its own is just theoretical potential. It can avoid hitting a tree or a pedestrian, but it needs to be taught what a pedestrian or a tree is

Furthermore, there is a huge variety of potential scenarios that could unfold whilst driving. “There’s a story about an autonomous vehicle on the road in Australia that got very confused when it came across a kangaroo, and couldn’t tell if it was an animal or a human. This training is an on-going, iterative process, as we have to account for every possible scenario.”

However, it’s not just highly autonomous vehicles that need to be trained. Barshack also provided an example whereby lane keep assist (a driver assistance system featured in most new vehicles on the market today) malfunctions.

“The lane assist feature will provide an alert when a driver crosses the lane marker, and will also pull the steering to correct the vehicle’s position,” she noted. “But if there’s a piece of garbage in the road and it happens to be yellow, it may or may not think I’m crossing a lane marker.”

All these random scenarios could have a big impact on the way that self-driving cars function, and in some cases, they could potentially cause the vehicle to react in a dangerous manner. Therefore, extensive training of the algorithms is important both in terms of the robustness and efficiency of the technology as well as its safety.

The bottleneck

OEMs have various timeframes in mind when it comes to the deployment of autonomous cars, but General Motors is hoping to lead the pack. The company has revealed plans to launch a commercial fleet of so-called robotaxis in 2019, and is leveraging the expertise of Cruise Automation to do so. Ford is preparing to have self-driving vehicles ready by 2021, and is currently testing its technology through partnerships with the likes of Lyft, Domino’s Pizza and Postmates, and various other vehicle manufacturers have referenced later dates.

“We’re seeing what I would call an arms race in the industry,” Barshack observed. “The company or companies that own the ability to power fully autonomous vehicles are going to win. I think it’s probably common knowledge in the automotive industry that if you’re not in the area yet, you are not preparing for the future.”

If we expect cars to be autonomous then they must be intelligent. They must be able to firstly identify a road sign, and then secondly be able to tell what that road sign says so as to react accordingly

Many companies already have permits to test their autonomous vehicle technology on public roads. A recent report by the California Department of Motor Vehicles (DMV) showed that the likes of Waymo, GM and Nissan self-driving vehicles have been successfully operating in autonomous mode.

“I think the computing capabilities are all there, such as the ability to provide 360-degree vision and to process high resolution images very rapidly,” Barshack stated. “We’re not really waiting for any additional technology. The bottleneck at this point is the training. It’s just a matter of how much manpower can we throw at this.”

Creating jobs

With the widespread rollout of autonomous vehicles on the horizon, some industry watchers have suggested that thousands if not millions of jobs could be threatened. Taxi drivers, for example, could be replaced by self-driving cars that are much safer and cheaper to use, and a similar scenario could happen with heavy-duty vehicles.

However, Barshack thinks that the job market will simply evolve: “There are already so many people getting worried about what taxi drivers or truckers will do, but the reality is that the job opportunities are shifting not disappearing.”

Self-driving vehicles will require people to train them, she added, and at least during the early stages of market introduction, specialist operators will be needed to sit in the vehicles and take control when needed. This is already the case for platoons of trucks that are currently being trialled on public roads today.

This article appeared in the Q1 2018 issue of Automotive Megatrends Magazine.