Autonomous driving and the giant leap from L1 to L5

Without a doubt, the automotive industry has painted an overly “rosy picture” of autonomous driving. The thought of being able to take your hands off the steering wheel, sit back and be liberated from the stressful not to mention dangerous chore of driving is seductive indeed.

The reality, however, is that fully autonomous driving is still quite far away. Industry insiders will rattle off a long list of roadblocks , from technology to safety issues, from business models to laws and regulations, which explain why autonomous driving is not just around the corner. Nonetheless the trend has been set, and manufacturers are racing each other to reach this ultimate goal, all striving to level up whether it appears to be viable or not. How this rocky road should be traveled, and how the roadblocks can be cleared, requires careful consideration.

Technically speaking, the challenge of realizing autonomous driving has always lain with the issue of scalability, since the ultimate goal of full driving autonomation is achieved stage by stage rather than in a single step. Therefore, the real challenge lies in building a scalable technical architecture that can be used throughout this long process of development, which can address the different computing power and safety requirements at each level of autonomous driving. In addition, as high-, middle- and low-end differentiated products emerge in this process to adapt to the needs of different user markets, this scalable architecture must also be able to quickly incorporate and monetize new technology.

Levels of autonomous driving

To fully grasp the enormity of this challenge, we must first examine how autonomous driving levels are defined. According to the definition given by SAE International, autonomous driving is divided into five levels, from L1 to L5, that respectively refer to driver assistance, partial driving automation, conditional driving automation, high driving automation, and full driving automation.

Figure 1 Illustration of autonomous driving levels (Image source: NXP)

Based on the capabilities of these five levels, which are described in Figure 1, it is plain to see that these levels are defined according to the degree of driving control. In other words, the lower the level of autonomous driving, the more control the driver has over the vehicle. For example, L1 constitutes several aspects such as adaptive cruise control, automatic braking and lane-keeping assist. In effect, these functions only allow the vehicle to perform automatic control of uni-directional acceleration or deceleration and not the actual act of steering. The driver still has absolute control over the vehicle, and is tasked with making correct judgments and decisions based on their personal observations of the environment. At L5, however, the vehicle is in a fully automated state that requires no driver intervention. In fact, the driver almost has no “say” whatsoever in the driving of the vehicle.

As you can see from these descriptions, a great "leap" occurs between L3 and L4. If the autonomous driving system from L1 to L3 is still a driver-oriented product that’s based on manual control of the vehicle, then at L4 and L5 the car is basically equivalent to a robot that is almost always disconnected from human intervention and operates independently. In truth, no matter how autonomous driving is “sold” to us, it is in its current form nothing more than ADAS. Only at L4 and L5 do you truly enter the realm of full driving automation.

When you take into consideration the giant leap between L1 and L5, the scalability of technical architecture seems even more challenging.

Scalable technical architecture

To even begin to overcome this challenge, we must first obtain an in-depth understanding of what it constitutes. Currently, industry developments can be divided into two areas: perception and modeling, and safe computing.

Specifically, perception and modeling aim to perform feature extraction, classification, recognition, and tracking of data. This data is obtained from vehicle sensors that gather information such as what the target is, the XYZ coordinate position of the target, and the speed and angle of the target's movement, and output a network diagram. The output of the perception and modeling domain can then be used as the input of the safe computing domain. What safe computing needs to achieve is to integrate the network diagram of the target with environmental information, plan the best route, and dynamically predict possible changes in the next few seconds. The output of the calculation result is the two control signals of vehicle acceleration and deceleration and steering. This calculation process is repeated to form a coherent automatic driving behavior.

As the two areas of perception and modeling and safe computing have different functions, their specific technical requirements also differ. These differences are reflected primarily in their functional safety and computing efficiency.

For perception and modeling, since the front-end input comes from multiple sensors, including camera lenses, mmWave radars and laser radars, at least two of these three types of sensors must meet the requirements of comprehensive and accurate data acquisition to adapt to complex application scenarios. Just one of the many sensors in the perception and modeling system needs to meet the functional safety requirements of ASIL-B for the system to achieve ASIL-D functional safety level as a whole. In terms of computing power, fixed-point computing can be utilized to meet the majority of requirements for sensing and modeling data processing.

Safe computing, however, is a very different story. Following sensor fusion, since there is no data diversity or redundancy, the safe computing processor must single-handedly meet all ASIL-D functional safety requirements. And due to the system’s high computational complexity, fixed-point computing and floating-point computing must be used concurrently, since floating-point computing is utilized mainly for vector and linear algebra acceleration. Furthermore, from the perspective of security, neural networks are incompetent because they cannot backtrack. Since deterministic algorithms must be used, these computational efficiency requirements require the support of a corresponding computing architecture.

Imagine, if you can, a single computing architecture that completes the two tasks of perception and modeling, and safe computing, at the same time. Obviously, this would be uneconomical and limit flexibility. For example, to expand the number or type of sensors, you would be required to replace the entire processor structure. Therefore, one way to design a truly scaleable architecture would be two use two different processor chips – one for perception and modeling, and one for safe computing, thus facilitating subsequent system expansion and upgrades. This is how NXP, for instance, is designing its products. 

Figure 2 NXP's scalable technology architecture for autonomous driving (Image source: NXP)

For perception and modeling, NXP has launched two processors, S32V for visual perception and modeling, and S32R for radar perception applications, processing data from photographic lenses and millimeter wave radars. There is ample computing power for them to serve as auxiliary processors that improve the overall safety of the entire system. This type of architecture is fully competent in ADAS applications at L1 and L2 levels.

For L3 to L4 level systems, S32V and S32R processors can be added accordingly if additional camera lenses and radars are needed. At the same time, as safe computing is already indispensable at this level of autonomous driving, NXP utilizes the S32G and Layerscape LX2 to undertake this task. S32G is used as a safety processor and automotive peripheral interface processing equipment, while LX2 is used to improve computing performance and to support high bandwidth and network transmission functions.

With this architecture, L4 and L5 levels can be achieved by continuing to add front-end perception and modeling processors as well as third-party accelerators to the system, thereby greatly enhancing the flexibility of system development.

Figure 3 NXP's self-driving technology architecture is flexible and scalable (Image source: NXP)

With these approaches, it is clear that one single architecture can meet the technical requirements of all autonomous driving levels from L1 to L5, enabling developers to pursue research and development for current market needs while pushing the limits of technology far into the future. Without a doubt, sooner or later, the “rosy picture” of fully autonomous driving will become the reality of today.