Succeeding with the automated driving journey through AI

For years, automotive original equipment manufacturers (OEMs) have been working towards realizing the dream of self-driving cars that travel without interruption on public roads, carrying their passengers in comfort and safety to their destinations.

There have been major strides towards this new reality. Many automotive companies are already embedding software with varying levels of autonomous capability into their vehicles, although only a handful currently have made such features controllable by the driver.

What’s holding things up? Availability of the technology needed for autonomous cars is not usually a problem – indeed, the technology is already installed in many cars currently on the road. A more significant barrier to progress in this area is posed by issues around the regulation and interoperability of the technologies.

It’s also well known that some companies have run into serious trouble after making autonomous driving functionality accessible: There have been injuries and even fatalities among drivers and pedestrians. Terrible outcomes such as these further complicate the legal issues surrounding the development and deployment of the technology, potentially setting back the rate of development by years. So how can we safely accelerate towards an autonomous driving future?

It's generally agreed that there are five levels of autonomous driving technology, with level 1 representing the inclusion of some automatic features, such as collision detection or lane departure warnings, and level 5 representing full automation, where no driver is needed.

For mainstream OEMs at least, the current focus is on the lower levels, with the term advanced driver assistance systems (ADAS) often being heard more than “autonomous vehicles.”

Currently, even the market leaders are just at level 3, and only features from levels 1 or 2 are legal for use on public roads. At those relatively modest levels, there are already important technology challenges. For example, “autopilot” mode requires the driver to be awake and alert, ready to manually intervene by steering and braking when required. This in itself requires hugely sophisticated technology, with the vehicle not just sensing conditions and events on the road, but also monitoring what the driver is doing.

Level 3 is where things get really tricky, and not just because the technology is exponentially more complex. In addition, every technology component at every step needs to be developed so that it can behave correctly in every situation. Then it needs to be independently validated and assured for its effective and safe use. Many of these difficulties can best be understood in terms of data. 

Higher levels of automation depend on a vehicle perceiving external obstacles via its sensors and making AI-enabled judgments based on distance and time, in a similar way to a human driver. For the vehicle to achieve this level of intelligence, a large amount of data needs to be collected and processed into a format that can drive those decisions. Some of this data will come from high-quality sensors and cameras attached to the vehicle itself, but clearly no vehicle could “experience” every possible situation. Crowdsourced data – for example, from other vehicles – is also required. 

This will soon add up to a huge volume of data. Big data techniques can help with the management of that data, but for it to be useful to the autonomous vehicle it has to be processed. Sensor and camera data must be converted into 2D and 3D exterior views, labeled to help AI models understand it (for example tagging cars, lorries, buildings, pedestrians, and so on in an image). This enriched data must then be turned into scenarios that can form the basis for automatic decisions and judgments.

Until now, this labeling and other enrichment of data has required significant effort by humans. Today, however, data volumes are becoming too great for human engineers to process efficiently. Fortunately, AI can help to streamline these processes too, creating what are essentially automated, crowdsourced truths. 

Increasingly, this type of automation will be critical to ensure that vehicles are equipped with scenario intelligence that is accurate and free of the redundancies that can make it harder to use.

Research suggests that large automotive OEMs can boost their operating profits by up to 16% by deploying AI at scale, provided they focus their AI investments in the right place. To realize this advantage in the autonomous driving arena, companies need to build or acquire competencies in a number of specific areas, for example: 

Intelligence: OEMs must build a detailed understanding of the role played by intelligence at all levels of autonomy and make sure that they can obtain, validate, verify, and standardize that intelligence.

Data: As we have seen, higher levels of autonomous driving functionality require computation and decision-making around large numbers of situations and decisions, and these depend on access to intelligence derived from large volumes of crowdsourced data. It is important to be able to manage and process that data effectively. AI and data competencies together are required to enable all the data to be stored, annotated, visualized, analyzed, and shared effectively.

Connectivity: Autonomous driving can be facilitated by better connectivity, which enables many aspects of the vehicle to be updated over the air (OTA). Coupled with the provision of functionality in software rather than hardware wherever possible, this helps to ensure that all vehicles are benefiting from the latest intelligence and technology.

Communication: 5G will be pivotal to enabling autonomous driving on public roads. For autonomous driving to work, enormous volumes of data have to be transferred, whether we are looking at automatic vehicle-to-vehicle (V2V) communications or the vehicle-to-everything (V2X) communication required for the anticipated smart infrastructure for roads. By accelerating connection speed and reducing latency, 5G enables vehicles to communicate almost instantly with each other and a huge number of connected on-road assets and infrastructure.

Even with the requisite competencies available, challenges remain. For some newer automotive companies, technology testing and deployment has been in their DNA from the outset, with the vision of autonomous driving a clearly identified strategic goal. But for longer-established OEMs, the journey towards autonomous car technology is arguably more challenging. 
For one thing, as ADAS technology becomes more mainstream, consumers will expect new cars to come fitted with level 1 and 2 ADAS features as standard. This means, among other things, that manufacturers will need to retrofit these new technologies into existing vehicle ranges, and therefore new elements will need to be added to existing design, testing, and manufacturing processes.

There are more fundamental issues, too. Traditional OEMs cannot easily redesign their development processes from scratch to accommodate ADAS. In addition, few of these companies have the resources or time to conduct their own experiments and develop new ADAS implementation processes. 

One way to overcome the obstacles is to spin off smaller companies or units that can innovate fast because they are unhindered by legacy processes. Another is to collaborate with an ecosystem of partners such as suppliers, research institutes, and other industry players. OEMs will benefit from working with partners who can develop working proofs of concept on their behalf – something that can slash development time and budget, and free the OEMs from some regulatory burdens. Partners can also help with acquisition of the necessary data sets and supplement in-house competencies in the key areas we have identified. 

Importantly, working with a partner ecosystem can help with the move to Intelligent Industry by supplying the product engineering and data science capabilities that are needed to master data and deploy technology at scale. They can also co-develop, or validate and verify, autonomous systems and technologies. It will accelerate the process of getting automated vehicles onto the road safely and successfully.