The 2020 Autonomous Vehicle Technology Report, collaboratively written by experts from across the automotive field, describes the current state-of-the-art of self-driving technology. Ultimately, the report found that important innovation is happening on all levels, and that specifically better software and reduced costs are key elements towards affordable autonomous vehicles capable of driving in all circumstances.
Initiated by the Wevolver platform after noticing a surge of interest in the topic, the report was written by and for the engineering community to create the first comprehensive understanding of the current cutting edge in the field
Autonomous vehicle technology covers many areas. This required a uniquely collaborative approach to the report's writing, reflective of complex and diverse nature of the field: A call out in the Wevolver community yielded an initial core group of writers that created a rough draft. Afterwards, the manuscript went through dozens of iterations and feedback sessions by experts and members of the engineering community.
These experts reflect a wide range of voices: From Maxime Flament, Chief Technology Officer, 5G Automotive Association, to Mark A. Crawford Jr, Chief Engineer for Autonomous Driving Systems at Great Wall Motor Co, they represent 11 countries, and bring together the views of startups, large enterprises, non-profits, and academic research labs.
Main conclusions from this group are:
- Fully autonomous systems need to outperform humans significantly to establish enough trust for acceptance by the general public.
- Software is the most challenging part of autonomous driving systems.
- All types of sensors (Cameras, RADAR, Ultrasound, etc.) are undergoing development, pushed by the demands of vehicle manufacturers. LIDAR sees the most innovation, as it’s moving away from the traditional, relatively bulky and costly mechanical scanning systems that are known from the iconic domes on top of self-driving cars.
- The increasing availability of alternative Global Navigation Satellite Systems beyond the American GPS, such as the European Galileo, enables to achieve orders of magnitude higher accuracy in positioning
- Most in the industry express High Definition maps to be a necessity for high levels of autonomy, though some notable exceptions, including Tesla and MIT researchers, focus on a 'map-less' approach, and experts point to the need for self-driving vehicles to become less reliant on existing maps.
- Deployment of 5G networks is seen by most experts as essential for successful implementation of autonomous vehicles.
- Often overlooked, internal and external user experience design is essential to be able to introduce self-driving cars and other road vehicles on a large scale.
Initially aimed at 3500 words, covering the topic completely turned out to require the final report containing over 16,000+ words of content, formatted into a freely downloadable PDF by the award-winning book designers from Bureau Merkwaardig. It's available for engineers who need to deepen their knowledge, for leaders who want to grasp the technological challenges and breakthroughs, and for all who are interested in learning how many fascinating technologies come together to create the mobility innovations that change our future.
Below you can find a summary of the report. The full report is available here.
The guide to understanding the state of the art in hardware & software for self-driving vehicles.
Motorized transportation has changed the way we live. Autonomous vehicles are about to do so once more. This evolution of our transport — from horses and carriages, to cars, to driverless vehicles — has been driven by both technical innovation and socioeconomic factors.
At the start of the 2020s, the state of autonomous vehicles is such that they have achieved the ability to drive without human supervision and interference, albeit under strictly defined conditions. This so-called level 4, or high automation, has been reached among many unforeseen challenges for technology developers and scaled back projections.
“It’s been an enormously difficult, complicated slog, and it’s far more complicated and involved than we thought it would be, but it is a huge deal.” — Nathaniel Fairfield, distinguished software engineer and leader of the ‘behavior team’ at Waymo, December 2019
No technology is yet capable of Level 5, full automation, and some experts claim this level will never be achieved. The most automated personalvehicles on the market perform at level 2, where a human driver still needs to monitor and judge when to take over control, for example with Tesla’s Autopilot. One major challenge towards full autonomy is that the environment (including rules, culture, weather, etc.) greatly influences the level of autonomy that vehicles can safely achieve, and performance in e.g. sunny California, USA, cannot easily be extrapolated to different parts of the world.
Beyond individual personal transportation, other areas in which autonomous vehicles will be deployed include public transportation, delivery & cargo, and specialty vehicles for farming and mining. And while all applications come with their own specific requirements, the vehicles all need to sense their environment, process input and make decisions, and subsequently take action.
Generally, a mixture of passive (cameras) and active (e.g. RADAR) sensors is used to sense the environment. Of all perception sensors, LIDAR is seen by most in the industry as a necessary element. Some are going against this conventional wisdom, including Tesla (relying on cameras RADAR, and ultrasound), Nissan, and Wayve (relying on cameras only).
These sensors are all undergoing technological development to improve their performance and increase efficiency. LIDAR sees the most innovation, as it’s moving away from the traditional, relatively bulky and costly mechanical scanning systems. Newer solutions include microelectromechanical mirrors (MEMS), and systems that do not use any mechanical parts; solid-state LIDAR, sometimes dubbed ‘LIDAR-on-a-chip.’
For higher-level path planning (determining a route to reach a destination), different Global Navigation Satellite Systems beyond the American GPS have become available. By leveraging multiple satellite systems, augmentation techniques and additional sensors to aid in positioning, sub-centimeter accuracy for positioning can be achieved.
Another essential source of information for many current autonomous vehicles are high definition maps that represent the world’s detailed features with an accuracy of a decimeter or less. In contrast, some companies, including Tesla and Apple, envision a map-less approach.
For the whole process of simultaneously mapping the environment while keeping track of location (SLAM), combining data from multiple sources (sensor fusion), path planning and motion control two different AI approaches are generally used:
- Sequentially, where the problem is decomposed into a pipeline with specific software for each step. This is the traditional, and most common approach.
- An End-to-End (e2e) solution based on deep learning. End-to-End learning increasingly gets interest as a potential solution because of recent breakthroughs in the field of deep learning.
For either architectural approach, various types of machine learning algorithms are currently being used: Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN) and Deep Reinforcement Learning (DRL) are the most common. These methods don’t necessarily sit in isolation and some companies rely on hybrid forms to increase accuracy and reduce computational demands.
In terms of processors, most AV companies rely on GPU-accelerated processing. However, increasingly different solutions are becoming available, such as Tensor Processing Units (TPU) that are developed around the core workload of deep learning algorithms. More electronics, greater complexity, and increasing performance demands are met by semiconductor innovations that include smaller components and the use of novel materials like Gallium Nitride instead of silicon. Engineers also face questions about how much to distribute or centralize vehicles’ electrical architecture.
To increase the available data for autonomous driving systems to act upon and increase safety, vehicles need to share information with other road participants, traffic infrastructure, and the cloud.
For this ‘Vehicle-to-Everything’ (V2X) communication, two major networking technologies can be chosen:
- Dedicated short-range communication (DSRC), based on a WiFi standard,
- Cellular V2X (C-V2X), which for AV applications needs to be based on 5G.
At the moment both DSRC and C-V2X are going through enhancements. The question whether DSRC or C-V2X is the best choice is a subject of debate. Due to its rapid progress and performance, the latter is increasingly preferred, and experts express that DSRC won’t sufficiently support some key AV features.
In parallel with technological development, user experience design is an important factor for autonomous vehicles. For lower level automated vehicles, where humans at times have to take control and drive, mode confusion can arise when the state of the vehicle is unclear, e.g. whether autonomous driving is active or not.
Other key challenges for user experience design are trust-building and communicating the intentions of self-driving vehicles. Internally, for the passengers, human driver behavior is often emulated on displays. For external communication companies are researching displays with words or symbols to substitute the human interaction that people heavily rely on when participating in traffic.
Wevolver’s community of engineers has expressed a growing interest in autonomous vehicle technology, and hundreds of companies, from startups to established industry leaders, are investing heavily in the required improvements. Despite a reckoning with too optimistic expectations it’s expected we will see continuous innovation happening and autonomous vehicles will be an exciting field to follow and be involved in.
“The corner cases involving bad weather, poor infrastructure, and chaotic road conditions are proving to be tremendously challenging. Significant improvements are still required in the efficacy and cost efficiency of the existing sensors. New sensors, like thermal, will be needed which have the ability to see at night and in inclement weather. Similarly, AI computing must become more efficient as measured by meaningful operations (e.g., frames or inferences) per watt or per dollar.” — Drue Freeman, CEO of the Association for Corporate Growth, Silicon Valley, and former Sr. Vice President of Global Automotive Sales & Marketing for NXP Semiconductors, December 2019
The full report is available here, for reading online, and as a downloadable PDF.