August 4, 2021 | Melanie Stone
Autonomy is here. While still a few years away from deploying full autonomy across all transportation, autonomy is currently robust — particularly across industrial use cases where Cyngn operates. In recent years, we have seen the development of autonomous vehicles continue. This can be attributed to the notable changes in the self-driving space that either didn’t exist five years ago, or have improved by many multiples in that time. These technological advancements in various domains have allowed the industry to increase capabilities for AV with higher levels of quality, quantity, locality, and efficiency.
In this article, we’ll explore the five massive transformations that have come to autonomous vehicles, and their positive impacts on the autonomy landscape. These transformations include new computational methods, improved sensors, advanced computing power, an open ecosystem, and 5G.
1. New Computational Methods:
Autonomy can be described as a super-set of hardware, software platforms, and various other tools. Accordingly, in order to discuss the recent improvements to computational methods, we must first look at how AI-based learning methods are being used in different aspects of autonomy today.
AI-based learning methods have significantly developed and optimized in the past five years. Autonomy begins with perception. A vehicle’s perception stack allows it to perceive the world around it. These improvements go beyond just the perception stack and are present in the entire system. The ability to perceive, classify, track, and predict objects is the place where learning-based messages shine, opening a door that didn’t exist several years ago. These advanced algorithms have enabled greater accuracy by improving awareness of the surrounding environment, which renders enhanced safety.
According to Cyngn’s Head of Autonomy, Biao Ma, many algorithms have recently been developed in autonomy across the stack. These algorithms form AI models that “replicate a decision process to enable automation and understanding”. This includes the architecture for a perception stack, as discussed, or say, the object direction classification stack. “If you compare these state of the art models today to not even five or three years ago,” says Ma, “there’s a significant improvement of accuracy, latency, and the efficient utilization of the compute unit.”
Data sets for these AI models have also transformed. There are not only a big variety of open data sets today, but those in 2021 are better registered, calibrated, and synchronized. By having an advanced way to synchronize sensors, it allows for a set of data that is a hundred times bigger at a much higher quality. Consider, Kitti Dataset, which was popular around ten years ago. This data set had around 7,000 frames with each labeled with what each object it is. Now, a popular dataset has about a hundred times that many frames.
Improved computational methods have also helped us combat the challenge of corner cases that vehicles may experience while driving in the autonomous space. Corner cases will exist for many years, so advanced computation methods are very important in autonomy. If computation is not done in an effective and scalable way, solving these long-tail problems becomes very difficult. Therefore, researchers are developing new computational methods and platforms to do three key things in an effective way (a) capture and create scenarios, (b) require and provide automatic grading, and (c) manage data pipelines to trigger computation and effectively store data.
2. Improved Sensors:
There has also been a significant improvement in the distance resolution of sensors in recent years, particularly LIDAR, a light detection and ranging technology. The number of beams in these units have improved from about 40 lines to more than 300 lines today. Furthermore, the range of these units has gone from about 25 meters of perception to more than 100 meters. This plays a vital role in AV because sensors help vehicles collect information that allow them to capture and react to the given environment. With more beams and greater distances, LIDAR can better create 3D representations of detected objects and surroundings, and expand a vehicle’s previously narrow field of view.
Lastly, the signal-to-noise ratio (SNR) has seen significant development by reducing the signal’s noise. SNR is the ratio of signal power to noise power, meaning as noise decreases, receivers can better decipher the signal. The SNR has reduced by almost 50% in the past five years. “This changes the nature of the work of perception in LIDAR,” says Ma. This advancement of sensor technologies in the last five years can also be seen in radar, cameras, and other sensors beyond LIDAR.
3. Advanced Computing Power:
Improved GPU computation power and video memory size has also benefited developers of autonomous driving. GPU computational power has increased threefold. “If you look at a learning-based model, its usefulness is bounded by how fast the model can go'' and an increase in speed facilitates the ability for different structures of the learning-based model, says Ma. We typically expect a learning-based model to go beyond 10 or 20 hertz, but if it is running less than this at two or five hertz, then it is too slow. With the latest generation of perception, models are focusing more on using the latest generation of GPU to increase efficiency and speed. This speed allows for greater opportunity for different structures of the learning-based model.
Another aspect is the VRAM or “video random access memory”. If a model is too big, it means that it cannot fit in the GPU. A RAM’s size changes the size of the model you can fit into a GPU, and these recent improvements have led to faster and bigger models. This changes the variety and richness of the learning-based model engineers can choose from. Specifically, later generations have more CUDA cores, which are parallel processors that help process that data that comes in and out of the GPU. They could also utilize Tensor cores, which can be leveraged from an open platform known as Pytroch or Tensorflow to create optimization. While before things were either too slow or unable to fit the model in the GPU, this is no longer the case.
4. Open Ecosystem:
The rise of the open-source ecosystem has also altered the AV landscape and allowed engineers to work cross-functionally. In the last five years there has been a list of open-source projects that have given those in the autonomous space the opportunity to learn from each other by being exposed to different angles and ways of implementations. Several years ago there weren't open platforms, open-source projects, or open tools like Pytroch or Tensorflow for developers to develop. Ma believes that “these are great tools and methods for young engineers looking into autonomy, who can utilize them as a starting point for their autonomous vehicle careers.” This has resulted in more and more data sets that are bigger in size and better quality.
5. 5G - New Communication Technology:
Lastly, the difference between 5G and 4G greatly affects AV development. While 4G is already thousands of times faster than previous generations, 5G presents a network that is a hundred times faster than 4G. The 4G network is not enough for these advanced technologies. The fifth-generation wireless technology allows for more than just higher speed and also results in greater locality and low latency. This expanded network improves reliability and generates greater capacity by providing more space at faster speeds.
Ma says that 5G changes the locality of how the architecture of autonomous systems could be designed. Simply stated, it changes where things can happen. Putting hardware in a vehicle, for example, has limits in terms of how much energy it can consume and how well it can compute. In the past, it’s been hard for developers to make edge devices 2X or 10X more powerful because there simply isn’t enough space or energy in a vehicle. However, 5G will allow the ability to put certain key components in places that are not bounded by power usage, bringing in many more opportunities for next steps in autonomy.
Ma further breaks this down into two key numbers related to 5G: the reduction of latency (how fast something is sent and delivered) and throughput (how fast you can get data and how big a load is sent). First, 5G is 1 or 2 milliseconds in terms of latency, which is more than ten times faster than 4G; this changes where these essential and time-sensitive computations happen. Second, 5G gets us beyond a gigabit per second of data transfer, meaning a significant load of data can get to the vehicle. These two components combined can lead to “a significant change in where things could happen and where it is more optimized”.
Consider the example of a light pole in regards to the computation of autonomy. With 5G, there are reduced bandwidth limitations and developers do not need to consider whether it’s too late to deliver that information to a given vehicle or not. By increasing speed, 5G can make it so various AV computations don’t have to live right inside of the vehicle.
Just in these past five years, the autonomous vehicle sector has seen these five massive transformations- new computational methods, improved sensors, advanced computing power, an open ecosystem, and 5G. Yet the AV sector will only continue to advance as we expect to see 5-or-10x improvements to autonomy in the following years. This is just the starting point. It begs the question, what will five years from now look like?