If we’ve learned anything about autonomous vehicles (AVs), we have learned that they are far from flawless. This is normal for any new technology that launches before enough “on-the-road” experience can be benchmarked against for the inevitable technology refinements.

The industry knows this and is moving forward with a new set of lessons learned.

Many of these refinements are in areas of more effectively collecting and processing data for navigation, communications and predictive analytics. 

“Achieving the same 99.99999 percent reliability of a human driver is the biggest challenge,” said Neil Cawse, CEO and founder of Geotab, which provides fleet tracking and vehicle connectivity software. “It’s still too hard for an autonomous vehicle to recognize and validate every bizarre situation that might be thrown at a human and deal with it correctly.”

One challenge is being able to recognize and track the acceleration of another vehicle while the AV is in motion.

“From a technology standpoint, acceleration tracking involves accurately monitoring your vehicle separately from other vehicles on the road and then tracking its change in speed over time to calculate its acceleration,” said Cawse.

From the AV's standpoint, acceleration represents just how quickly the vehicle’s speed is changing. As an example, braking is registered as a negative acceleration.

“Knowing acceleration is very important because speed is not enough if you want to accurately predict where a vehicle will be sometime in the future,” said Cawse. “Acceleration is also a critical element when a potential accident situation emerges. Because the situation can happen in a split second, acceleration measurements must happen very quickly and be accurate.”


What if the technology fails?

“This is the problem with autonomous vehicles,” said Cawse. “If they are not reliable enough and humans need to watch them, then it’s better the human drives to force attention. There are techniques such as using different neural net systems to check each other, thus significantly improving reliability, but if multiple neural nets still can’t handle a new unique situation, then it doesn’t help.”

Because neural nets are mathematically designed to simulate the human mind, which in itself can be unpredictable, we can’t always be sure that neural net analyses will be accurate. 

“A neural network can be designed to provide a measure of its own confidence in a categorization, but the complexity of the mathematical calculations involved means it’s not straightforward to take the network apart to understand how it makes its decisions,” said Will Knight, MIT Technology Review’s senior editor for artificial intelligence. “This can make unintended behavior hard to predict; and if failure does occur, it can be difficult to explain why. If a system misrecognizes an object in a photo, for instance, it may be hard (though not impossible) to know what feature of the image led to the error.” 

New chipsets with higher processing capabilities will improve the performance of neural network technology, as well as other supporting technologies like GPS that are used by autonomous vehicles. “This is because autonomous vehicles require high levels of power and processing to do neural net calculations,” said Cawse. “These calculations are best performed by GPU processors which consist of up to thousands of small processing units which do the work in parallel.”

The improved chipsets for neural nets will be complemented by improved chipsets for vehicle GPS systems. The new GPS chips will improve vehicle performance by removing reflections from GPS signals that can be distractors in difficult situations. These chipsets will also extend support for the plethora of tracking systems used in different countries, such as Galileo, GLONASS and BaiDou. 

These chips and processing improvements ultimately bring the discussion around to data collection and how we are going to manage the tsunami of data that will be bombarding vehicle control systems from myriad data collection points.

“There is a balance between how long you hold onto the data to decide what should be discarded,” said Cawse. “This is a trade-off because the more regularly you send data, the more accurately the server knows the location of where the vehicle is. There are techniques, such as sending heading and speed, that allow the server to calculate the position without receiving new data.”

It is likely more empirical results will be needed before “comfort points” on data collection can be arrived at.

Meanwhile, the automotive and technology sectors are moving forward on new ways for autonomous vehicles to cross-communicate with each other.

“There is V2X (vehicle to everything) communication, which includes V2V (vehicle to vehicle) and other standards that allow vehicles to communicate with infrastructures like traffic lights, or services on the Internet or other vehicles to control traffic and safety,” said Cawse. “Traffic flow rate data is also available today from different sources, allowing live monitoring of flows.”

What can we expect from autonomous vehicles in the next few years?

“We have seen what GPS and mobile with some smart servers did to the taxi industry with Uber,” said Cawse. “Telematics and the connected car have the same capacity to turn other industries, like home delivery, car share, and even the traffic that the general public has to endure, completely on their head. There is a lot more innovation that is going to occur in the connected car and telematics space.”