Editor's Note: The following is a guest post from Mohammad Musa, CEO of Deepen AI.
In a November interview with Recode, Tesla founder Elon Musk predicted that his company could have self-driving cars on the road as early as this year. However, as someone working within the industry, I predict that this will be another "smoke and mirrors" milestone which fails to meet expectations.
Elon Musk might be able to send Teslas into space, but we are still to see a successful full public launch of a driverless fleet by any of the largest players in the industry. The closest contender, the mega-hyped launch by Alphabet’s Waymo in Phoenix, AZ last November, has been referred to as "underwhelming." Rather than being open to public use, the launch was limited to a small group of around 400 early riders. The majority of commercial AVs still feature a trained human driver behind the wheel and reportedly still struggle with left turns when being controlled by AI.
To make matters worse, since autonomous fleets started live testing on public roads, a small number of high profile accidents have reduced trust within the general public. If regulators get spooked and start setting unreasonable requirements on all testing and deployment of AVs, this could slow down the widespread adoption of the technology even further.
So, considering that Waymo, the largest, and most well-funded driverless vehicle offering, is still far from its final destination, how can industry players finally move past automated vehicle roadblocks and achieve widespread adoption?
1. Using the right data to speed up testing
One of the main challenges facing manufacturers of autonomous and driverless vehicle technology is the amount of time it takes to adequately test new systems. Autonomous systems are so complex that verifying safety requires a huge amount of simulation, which in turn equates to a huge amount of time and resources.
Currently, while manufacturers need to comply with Federal Motor Vehicle Safety Standards and certify that their vehicle is free of safety risks, to date there is no official legislation on testing requirements. But that's not to say that more strict regulations won't emerge. After all, the regular auto industry has strict testing regulations like Misra-C in place, and a number of state and federal laws have already been created to start bringing the industry into line.
In a 2016 article, U.S. scientist Gill Pratt estimates that to guarantee with 95% confidence that a driverless vehicle would be safer than using a human driver, it would require driving 8.8 billion miles. This would take a fleet of 100+ vehicles driving 24 hours a day, 225 years to complete.
This creates a kind of Catch-22 for manufacturers, as every time they make a major change, they need to do another round of testing to make sure the overall strength of the system has not been compromised.
Design changes that may seem insignificant, such as updating a sensor on a vehicle from version one to version two, affect the overall computing power of a vehicle and could ultimately put the whole system out of sync. Upgrading from a 32 laser beam LiDAR to a 64 laser beam LiDAR would increase the data usage, and could in turn severely affect compute capability. A minor design change like this would require a large amount of additional testing to validate a vehicle’s safety and reliability. After all, if a vehicle is traveling fast, in challenging conditions, a small change in computing loss could slow a system’s reactivity, increasing the chance of an accident.
There are an infinite amount of these types of scenarios that make validation and testing challenging. However, there is only one answer, and it lies in manufacturers doing a better job with dealing with their data.
The better companies understand their vehicles’ behavior and the data being collected, the better they will be able to analyze and predict any problems that could potentially happen down the line. Having a huge wealth of clean data to play with will also allow manufacturers to do a better job of simulating scenarios digitally using AI, to reduce the amount of live testing which needs to be done.
2. Gathering enough data to train deep learning AI
In a recent Smart Data Collective article, Dan Matthews argues that the battle for a top spot between industry giants will be ultimately determined by the company which can best use its data to maximize the efficiency of its AI.
To effectively train deep learning neural networks — the AI which controls vehicles’ perception in self-drive mode — manufacturers need access to an astonishing amount of data. To make real-time driving decisions, vehicles need consistent access to high-quality data which can be used at the edge, where sensors collect and analyze ever-changing road conditions. Autonomous vehicles also need to use extremely accurate predictive analytics to tell when an event is about to occur, such as when a pedestrian is getting ready to cross, or when a child runs onto the road.
The challenges are that these sensors and systems use an extremely high level of computing resources, and that deep learning neural networks are very sensitive to small changes in input data.
For example, in dim light, a stop sign could be interpreted as a speed limit sign by adding a few post-it notes or graffiti. This level of sensitivity to small changes in the input is very dangerous. Traditionally, deterministic software algorithms are not sensitive to changes like this in the input. Deep learning systems are a lot more flexible but don't have the same level of reliability.
As such, industry leaders have been gathering data to train algorithms for years, so as to be able to simulate any possible situation. In February 2018, Waymo announced it had logged 5 million miles of self-driving data with its Chrysler Pacifica minivans across 25 cities in 5 different states, adding to more than 5 billion miles of testing data collected via computer simulations. Uber has been collecting data from more than 300,000 vehicles driving in different climates and terrains all over the world.
However, while bigger players like Waymo, Uber, Tesla, and Cruise, are already well ahead in their data resources, gathering the necessary amount of data needed to train deep learning algorithms is much more challenging for smaller scale, up-and-coming mobility startups like Aurora, Zenuity, Autonomous Intelligent Driving, Pony AI, JingChi or Xiaopeng Motors.
Large original equipment manufacturers (OEMs) are increasingly moving into the world of autonomous vehicles too, facing similar challenges despite their large budgets. Volkswagen recently announced a budget of $50.2 billion to invest in electric, automated vehicles, joining other leading manufacturers like BMW, Daimler, Ford, and GM in the race to make driverless vehicles a mainstream consumer product.
However, it could be argued that these investments are not being made wisely. With other major players like Tesla, Uber, and Waymo so far ahead in terms of AI development and the data resources needed to train AI, it might make sense for OEMs to instead start partnering with software and AI companies, rather than developing in-house, to better help them traverse this new technological wave. Otherwise, it may be years before we see mainstream car brands with ‘level 4’ driverless abilities.
Another option would be for smaller players to band together to gain ground on the industry leaders who have until now left them in the dust. Collaborative platforms such as Apollo by Baidu, accelerate the development, testing, and deployment of autonomous vehicles technology. As more partners use the platform, more and more data becomes available, allowing all partners to fine-tune their systems quicker than if they were working alone.
3. Making profit margins larger
One of the main reasons that experts predict that autonomous vehicles are more likely to succeed for service providers — such as taxi services or delivery vehicles — rather than personal owners, is due to the extremely high manufacture and R&D costs involved.
While Advanced Driver Assistance Systems (ADAS) will be rolled out more and more in personal vehicles, level 5 full Autonomous capability will be too expensive for the average consumer, even after the tax credits and incentives being used as a carrot on a stick by brands like Tesla.
There are already razor-thin profit margins in the auto industry for manufacturers and vendors. For sedans, the typical margin is very low, at roughly 3-4%, and the only companies really making considerable profits are luxury car brands — but even these only have around 7% margin. With even the largest OEMs like BMW struggling to maintain profit levels amid heavy investment in new emerging technologies such as electric and self-driving vehicles, this puts a high bar of entry for newer, less well-funded players in the mobility sphere.
When hardware cost is so high, and manufacturers are not selling large volumes of units, companies using driverless fleets have to be extremely inventive in how they can get the most ROI from their vehicles.
Boosting margins will require reducing downtime by ensuring vehicles frequent high traffic areas with lots of demand and by speeding up processes like cleaning and refueling via robots and automation. It will also mean using new technology to maximize the customer experience, by adding in-vehicle entertainment, and AI systems to monitor the condition of vehicles to ensure cleanliness and safety.
So, while it may seem that driverless vehicles have been waiting on the starting line for an eternity now, I think it is unlikely that Musk’s predictions that private fleets of fully automated vehicles will hit the streets this year will come true. However, if big and small players alike can stay on top of their mapping and vehicle data, and smaller players start working more closely together, we could see some live public launches of game-changing services happening in the not too distant future.
And we can only hope that these cars will not require human drivers, or have problems making left turns.