Autonomous driving: the current state of play
The “DARPA Grand Challenge”, first organized by the US Dept. of Defense in 2004, gave autonomous driving a big boost. Autonomous vehicles (AVs) represent the most important disruption since we moved away from horse-pulled carriages and will revolutionize mobility as we now know it. AVs have the potential to drastically reduce casualty — humans are the cause of about 90% of deaths on the road. AVs will foster the development of mobility as a service, which will allow more people to be mobile, in particular those who cannot afford a vehicle or can no longer drive. In addition, AVs will allow “drivers” to make much better use of their time, especially while cruising on highways or stuck traffic jams. There are also economic benefits for the emergence of AVs, from driver costs to fuel economy. But the impact of AVs will not be only positive; a massive deployment in taxi and truck fleets will have a devastating social impact; this will have to be managed carefully.
What is the current state of play in autonomous driving? Who are the key players? What are some of key technical and regulatory challenges? In what settings are experimented AVs likely to be deployed first?
Where are we today with autonomous driving
Autonomous driving is progressively coming to reality thanks to new sensor technology, HD mapping as well as powerful computers for analytics, sensor fusion, deep learning and path planning. This will allow to safely address the 3 key building blocks of autonomous driving systems, namely sensing (the driving scene), mapping (finding the vehicle’s precise location) and driving policy.
Several carmakers have announced dates for the possible introduction of Level 4 autonomous vehicles — at least from a technical standpoint. Tesla, which has a head-start with Autopilot, talks about of 2019. For Ford, Nissan, Volkswagen/Audi, BMW, Volvo or PSA, it will be 2020 or 2021 with an initial focus on mobility services and/or fleets. Yet, the path is still long and steep. Industry participants introduce increasingly sophisticated advanced driver assistance systems (ADAS) to prepare for full autonomy. But there are still tremendous technical roadblocks on the way to Level 4 (or 5) autonomy. These are the ability to essentially understand, 3D-model and predict the driving scene, find the vehicle’s exact location (within centimeters) and finally safely negotiate its way among road users and obstacles.
Experimentation is paramount to develop systems and ensure safety
Companies have learned a tremendous amount and continue to do so from experimentation. Google has accumulated over 2 million miles of autonomous driving since 2009. Delphi had crossed the continental US. Most carmakers and several major Tier 1 suppliers have been testing AV on open roads. Pioneer companies EasyMile, Navya and Local Motors are already testing their Level 5, 10-12 passengers shuttles in over 10 countries, in cities like Paris, Las Vegas, Dubai or at industrial sites.
Ridesharing services have also been experimenting with autonomous driving — this is of strategic importance as we will see later. Uber performed a live test in Pittsburg last year and is about to repeat the experiment in Arizona. Nutonomy organized a similar test in Singapore in 2016.
Trucks are potential beneficiaries of the current development. Mercedes Trucks has been experimenting autonomous driving for some time. Otto Trucks (now Uber) was founded in early 2016 to create add-on equipment for trucks to operate autonomously; they tested their solution on open roads. Peloton Technology pairs trucks via a cloud-based operation center, synchronizing acceleration and braking for both vehicles. According to the company, this yields a 4.5% fuel saving for the leading truck and a 10% saving for the following one.
Artificial intelligence is a core component
AI is at the heart of the development of autonomous driving systems, contributing mainly to object recognition and decision making protocols to quickly boost safety. This explains why a number of AI startups that have emerged to focus on automotive; they include Nuro, Drive.ai, PlusAI, or Argo AI in which Ford last month committed $1 bn over the next five years. Big tech players, such as Nvidia have already gained significant experience in the field, but reproducing our brain’s capacity to analyze situations and make the right driving decision every time may be more difficult that AI developers think.
Map-based or end-to-end AI-based autonomous driving
3D/HD maps allow for vehicle localization within centimeters by comparing the scene captured by sensors and that are stored on the map. Massive efforts are being made to progressively build 3D/HD maps of the road network — there are about 4 million miles of roads in the US alone. Some players drive dedicated mapping vehicles, some crowdsource the data from vehicles in service. Crowdsourcing will provide real time updates, allowing vehicles to know at all times where construction sites or potholes are located.
There is an emerging debate about the relative roles of 3D/HD maps vs AI assistance. UK’s FiveAI promotes an end-to-end autonomous driving solution with no 3D maps required, combining advanced computer vision and deep neural networks. Even if a pure AI-based autonomous driving solution may not be for the near term, progress made in this area will benefit map-based solutions and increase overall safety.
Who are the key players
Several carmakers took part in the “DARPA Grand Challenge” including Ford, Volkswagen and GM. Today, the leading carmakers are Tesla, Ford, the German companies, Volvo as well as Renault/Nissan not far behind. They invest heavily in in-house resources as well as in startups, as demonstrated by GM’s acquisition of Cruise for $1 bn in March 2016, of Ford’s very recent $1 bn commitment to Argo AI.
A few Tier 1 automotive suppliers are betting on the market, such as Valeo, Delphi, Continental, ZF/TRW, Bosch, Renesas or Autoliv. They progressively bring to market sensors (camera, LiDaR, radar) as well as data fusion hardware and software. But the integration of AI and heavy-duty computing power in the system requires new skills. This is why Bosch and ZF recently announced they will partner with Nvidia to integrate the chip maker’s Drive PX computer in order to complete their AV system.
The foreseeable growth in AV brings major tech companies into the game. Besides Waymo’s early move (following Google’s acquisition of 510 Systems), chip makers have been investing massively in AV, in particular Intel, Nvidia and Qualcomm. Nvidia will have its Level 3-capable Drive PX car computer ready by late 2017 and Level 4 in 2018. The most recent move is Intel’s $15 bn acquisition of vision specialist Mobileye this March. Intel’s preexisting 15% equity in map-making Here helps make Intel-Mobileye-Here a key piece in the AV chess game. Intel-Mobileye expects their turnkey system to cost about $5,000 by 2019.
Maps are critical a piece of the autonomous driving puzzle, at least for the time being. This explains why Daimler, Audi and BMW bought Here from Nokia for 2.8 bn € in 2015. Whereas Google or TomTom build their maps by driving dedicated vehicles, Tesla, Civil Maps — and soon Here — crowdsource their data. Mobileye plans on using the 15 million vehicles already fitted with its cameras to provide the data. It has also signed agreement with Volkswagen and BMW to install its data generation technology (REM) on their vehicles as from 2018, with the objective to support the creation and updating of HD maps; BMW and Mobileye will transfer the data to Here. It is interesting to note that China’s NavInfo recently became a shareholder of Here. All these players are putting their pieces into position on the chessboard.
Driving autonomy will be a matter of life or death in the long run for ridesharing companies like Uber, Lyft of Didi. This explains why Uber acquired Otto 6 months ago and is testing AVs. Not only will AVs offer significant benefits (no driver cost, fleet management…), their existence will lower the barrier to entry. In fact, carmakers have been busy building partnerships in the mobility-sharing space for this purpose. Speed is of the essence for ridesharing players!
A regulatory environment under construction
The “Geneva Convention on Road Traffic” (1949) did not account for autonomous driving! Country by country, state by state, authorities are testing the waters, but we are yet to see legislation that fully takes AV into account. Last September, the US Federal Government issued an Automated Vehicle Policy articulated around a 15-point safety assessment, which sets an overall framework that still makes space for innovation. California is going one step further: the state’s Dept. of Motor Vehicles just published a proposed regulation to establish a path for the testing and deployment of fully autonomous vehicles. The essential part of the regulation addresses the private use of fully autonomous vehicles on public roads.
Obviously, companies have been allowed to test their prototypes on open roads to prepare for full scale deployment. Several countries, states and cities around the world have granted permits for private companies to test AVs with trained “drivers”. In California alone, 27 companies are now authorized to test AVs; they include incumbent carmakers and system suppliers, new carmakers and pure tech players. Last month, PSA was granted a permit to operate AVs on French roads with untrained “drivers”. By the end of 2017, It will be possible to test vehicles without driver/passenger onboard on Californian roads. The yardstick is moving!
Level 5 shuttles will soon be part of the landscape being the first fully autonomous vehicles to be commercially available. Costing currently $200-250k and operating at up to 25 mph, they will first operate on well defined A-to-B-to-A routes or be used in confined environments, such as university and corporate campuses, industrial sites or residential communities. These vehicles will allow for automated mobility on-demand, supported by cloud-based fleet management and user interface services offered by companies such as BestMile, Renovo, Vulog or RideCell. As far as individual vehicles, carmakers have announced the availability of their Level 4 vehicles from 2019 to 2021. Whereas the deployment of these vehicles on highways is potentially within reach, it will be much more difficult in dense urban environments where understanding and predicting the driving scene is be much more of a challenge. Imagine trusting your AV around Paris’ Arc de Triomphe at rush hour!
Trucks will be a strong beneficiary of autonomous driving — except from a social standpoint. In the short term, the significant fuel savings provided by platooning should trigger massive take-up for solutions like the one proposed by Peloton Technology (see above). Whereas mining trucks have been operating autonomously for a couple of years already, other types of vehicles should also reap the benefit of progress made in automotive, such as forklifts or agricultural equipment.
Insurance and liability will be significantly impacted
The insurance sector will be heavily disrupted. Since humans are the cause of about 90% of deaths on the road — and probably a similar percentage of crashes in general — insurance premiums ought to reflect the benefits of autonomous driving. The NHTSA’s study that followed Tesla’s May 2016 fatal crash shows that the carmaker’s crash rate dropped almost 40% after Autosteer (part of the Autopilot systems) was released. Banking on this promising finding, US insurance company Root recently introduced a premium discount of about 10% for Tesla drivers for all miles driven while they detect Autopilot is being used. Another aspect of mobility insurance will be liability once we reach higher levels of autonomy. Who bears the responsibility for an accident? In the case of Level 3, how do you define whether the “driver” regained control fast enough?
And then what…
Mobility and safety will benefit immensely from autonomous vehicles. But will there still be space for driving pleasure? Toyota says they will not eliminate the option for the driver to take over. And how about BMW’s “driving pleasure”? A better future may be to still make cars dedicated to actual driving, possibly on dedicated roads, while we let computers get us from A to B as we work, sleep or watch a movie.
Also published on LinkedIn (