Autonomous Trucks: Self Driving Convoys are YEARS Away
SEE LAST PAGE OF THIS REPORT Paul Sagawa / Tejas Raut Dessai
FOR IMPORTANT DISCLOSURES 203.901.1633 /.553.9827
psagawa@ / firstname.lastname@example.org
October 24, 2017
Autonomous Trucks: Self Driving Convoys are YEARS Away
We believe that the commercial viability of self-driving long-haul trucks will trail well behind the adoption of robo-taxis offering on demand service to consumers. While autonomous trucks would have enormous economic benefits, and could be limited to less complex highway conditions, the trucking use case presents substantial challenges beyond passenger cars – e.g. more severe impact from weather, less precise handling, lower margins for error, and higher costs of failure. Trucking is also subject to strict driver and road-use regulations that will be further obstacles to developing and implementing fully autonomous long-haul trucking. Moreover, road testing for self-driving trucks is years and millions of miles behind autonomous car platforms, and the learning is largely not transferable. As such, in the next several years, autonomous controls will likely be used only to assist human drivers to improve safety (e.g. accident avoidance, “autopilot”) and efficiency (e.g. coordination of platooning), leaving the biggest benefits for a much later date. We remain bullish on a much faster timetable for autonomous on-demand personal transportation services within well-mapped geofenced territories, and expect use cases for larger vehicles running fixed routes at modest speeds (shuttle buses, regular delivery services, etc.) will be viable soon as well.
- Autonomous trucking is a huge opportunity. Driver wages and benefits are almost 40% of the operating costs of a long-haul truck – those costs would obviously be lower for autonomous operation, even with drivers needed for off-highway segments at either end. Aerodynamic platooning and electric motors could reduce the 25% of costs that go for fuel. Perhaps most importantly, autonomous trucks would not be subject to the 11-hours/day limits for drivers, enabling much higher equipment utilization and cutting delivery windows in half. In a $676B US market ($2.2T global), self-driving tech could save $211B in costs, resolve a growing shortage of drivers, create significant competitive advantage for early movers and potentially draw business away from rail and air transport alternatives.
- Highway driving is much easier than local. Proponents have focused on highway driving, where passenger car systems like TSLA’s Autopilot have been commercially available for many months. Ostensibly, human drivers could maneuver trucks to highway on ramps, with another driver taking over when the truck exits on the other end of the trip. In between, the highway driving task is much easier for an autonomous system – no sharp turns, no pedestrians, more predictable traffic flows, etc. – and constitutes most of the trip schedule.
- Trucks are much harder than cars. Trucks are 15 times the weight of a passenger car and take almost 70% longer to come to a stop. They are typically 3 times longer and 30% wider, with less than 20 inches of clearance on each side of a lane. They are much more difficult to maneuver, and are much more seriously affected by weather conditions. The consequences of an accident are much more likely to be fatal, and even minor incidents can yield expensive delays. Trucks are subject to inspection and must guard against theft – typically responsibilities of a human driver. Trucks must also follow extensive regulations – driver limits, changing highway restrictions, etc. – that do not apply to cars.
- Autonomous cars are many years and millions of miles ahead in testing. GOOGL began work on self-driving cars in 2008 and has logged >3.5M miles in a wide range of driving conditions, with billions of variations tested through simulation. Given unique challenges, data collected from cars is of small value for self-driving trucks. Specific testing on trucks is barely underway and will be expensive and time consuming, involving many miles of conducted fully loaded and equipped semis driving through the full panoply of highway conditions.
- Years from self-driving long-haul trucks. Uber’s Otto subsidiary may have successfully delivered a truck-load of beer across 120 miles of Nevada desert escorted by an armada of safety vehicles and a trained driver and engineer ready to intervene, but actual driverless trucks could be a decade away. First, the extensive systems development and testing must be completed. Second, state and Federal regulators must allow fully autonomous trucks to pull full loads without drivers aboard, and/or dramatically increase the daily hours a human driver can ride in a moving truck. Finally, trucking companies must embrace the technology – investing $23K per truck, building an organization and infrastructure for local handoffs, convincing insurers to cover the vehicles and their contents, etc. While autonomous driver assistance (e.g. collision avoidance, etc.) may be quickly available, self-driving will not.
- Geofenced robo-cabs almost here. Meanwhile, GOOGL’s Waymo has announced the imminent launch of on demand and totally driver-free robo-cab service for subscribers within well-mapped environs around Phoenix. With this, we believe GOOGL is many months ahead of would-be rivals, and could realistically begin mapping and testing in additional target markets for service launch before the end of the decade. Companies like Uber, GM, BIDU, and others with fleet-service, full autonomy aspirations will likely follow in geographic markets not initially addressed by Waymo. See our recent piece (http://www.ssrllc.com/publication/self-driving-cars-building-a-team-to-bring-taas-to-market/)
- Fixed route delivery and shuttles are viable now. While the focus has been on long haul, we believe the first viable commercial appearance of large autonomous vehicles will be for local use cases on fixed routes, such as shuttle buses or regular delivery circuits (e.g. mail delivery, commercial supplies, etc.). GOOGL’s first implementation of the technology was for the employee shuttles that operated on its own campus, and airport shuttles would be an obvious analog.
- Who Wins? While we believe that disruptive approaches to self-driving cars – on-demand fleets of robo-taxis sold as a subscription service – will win the day, favoring the ecosystems built by technology leaders (i.e. GOOGL), autonomous trucking is likely to favor incumbents. All the major players in the industry, truck makers and operators alike, have been energized, and we do not expect an outsider to use self-driving tech to disrupt competition. Rather, the tech will come from within or be bought from 3rd party systems vendors, and the early effective operators will use it to gain advantage over the stragglers. Still, for the reasons outlined in this piece, we expect adoption to be cautious and years behind passenger cars. In that, the tech winners are very difficult to call now.
Cars Before Trucks
GOOGL is suing Uber over terabytes of technical files that it alleges were stolen by former lead engineer Anthony Levandowski and subsequently transferred to the ride sharing firm when it acquired his recently launched company, Otto. A bit lost in the shuffle of the splashy legal battle is that Otto was the first of the mushrooming population of self-driving startups to focus on long-haul trucking. Recently, TSLA threw its hat into the ring, announcing its own electric autonomous truck project, while major truck makers – e.g. Daimler, Volvo, Peterbilt, etc. – have begun their own initiatives to add autonomous controls to their products. Unlike personal transport, where we believe disruptors offering on-demand autonomous robo-cabs as a subscription service will carry the day, we believe trucking will favor an incremental approach working with and within existing OEMs and trucking service companies.
These incremental driver assistance systems – such as accident avoidance, short-term autopilot, and multi-truck coordination for platooning – are built on similar platforms to full self-driving and have real benefits. Fuel spending, roughly 25% of a truck’s total operating costs, could be cut 5-10% with possible reductions to insurance costs (~6% of total) once the systems prove their ability to reduce the frequency and severity of accidents. Still, these savings alone won’t make the technologies a competitive hammer for trucking companies. Labor costs are 40% of trucking expenses. Eliminating the 11-hours per day limitation on drivers could halve that on a per mile basis, while obviously cutting delivery intervals for customers. Pulling the driver entirely would yield even greater benefit. This is the pot of gold at the end of the rainbow.
Unfortunately, developing fully autonomous systems will take many years for long-haul trucks, even if self-driving is limited to simpler highway segments. Driving trucks is much more difficult, and carries much higher risks with lower margin for error, than driving cars – the physics of the size and weight of vehicle dictate that. Maneuverability is poor, stopping distances are long, the effect of wind and weather are exaggerated, and the outcomes of accidents are more often fatal and always expensive. As such, license requirements for human drivers are much more exacting than for cars and the rules under which they must operate far more stringent. The same attention will be paid to autonomous driving systems.
Development and testing for trucks is also far behind self-driving cars. GOOGL has been test driving robo-taxis for nearly a decade, has iterated on its hardware designs multiple times, has 3.5M miles logged on public roads under a wide variety of conditions, and runs billions of miles in simulation each year. In contrast, Otto’s test truck has driven 120 miles without lane changes and buffered by escort vehicles on all sides. Because the physics of the vehicles are so different, development work for self-driving cars does not easily transfer to self-driving trucks – fully autonomous trucking is starting nearly from scratch.
We believe that it will many years before driverless trucks are viable and approved for commercial service, even restricted to highway driving. Driver assistance systems will continue to get better and cheaper, and will be implemented by truck operators as they become economically sensible. Meanwhile, robo-taxi service is coming soon – GOOGL’s Waymo promises driverless rides to subscribers in Phoenix by year-end. We expect the transportation-as-a-service model will be well established in many metro markets long before convoys of fully autonomous trucks are on the road. The real short-term opportunity for truck-sized autonomous vehicles is more likely in fixed route local deliveries and shuttle buses – perhaps coming soon to an airport near you.
Wouldn’t it Be Nice?
As it dawned on the tech world that self-driving cars might be possible far sooner than even science fiction authors had predicted, many began opining that, perhaps, self-driving trucks might be the better target market. Consumers might balk at the technology, particularly if it required unsightly sensor mountings and high prices. In contrast, one forward thinking trucking company, convinced of the economic savings and competitive advantage to be gained from the technology, could drive the whole market ahead (Exhibit 1).
Exh 1: Incentives for Autonomous Trucking Technology
It is well understood that highway driving is less complicated than local driving – tens of thousands of Tesla owners are already using that company’s Autopilot system to give them a break from full attention under normal highway conditions, adding steering, breaking, acceleration, and other common functions to the speed regulation of ordinary cruise control. There are no pedestrians, bicycles, cross traffic, parked cars, or other such obstacles to consider, and relatively few turns or need to interpret unusual driver behavior. Meanwhile, most trucking hours are spent on those highways. Without drivers, running in energy efficient platoons, and avoiding the accidents precipitated by tired, inattentive or poorly prepared drivers, autonomous trucks could be a game changer.
The savings would derive not just from removing the driver, which constitutes about 40% of industry costs today, but also from coordinating platoons of trucks driving together to cut the wind and save 5-10% of fuel expenses, and from sharply lower accident rates that would yield lower insurance premiums (Exhibit 2,3). Further competitive advantage would come from obviating the 11-hour driver restriction that keeps deliveries parked for more than half of each day, thus cutting delivery intervals in half while raising the capacity utilization of the trucks themselves. Of course, there would be some new costs – self-driving would likely stay restricted to highway conditions, so drivers would be needed to conduct the trucks between loading docks and highway on/off ramps and would swap in and out as the vehicles checked in at facilities just off the interstate road system. The makers of trucks and the companies that operate them are certainly intrigued.
Exh 2: Current Cost Per Mile – Regular Cars and Long-Haul Trucks
Exh 3: Cost Per Mile – Autonomous Cars vs. Autonomous Trucks
The Problems with Trucks
Highway driving may be a much easier problem for AI-based self-driving systems to navigate, but driving trucks is significantly more difficult than cars. Most of the issues derive from the size of the typical 18-wheel tractor-trailer (Exhibit 4). Trucks can be 80 feet long, more than 5 times the length of a typical passenger car, and 10 feet high. Trucks are typically 8.5 feet wide, leaving just 20 inches of clearance on either side of a lane, about half of the clearance for cars (Exhibit 5). Fully loaded trucks can weigh up to 40 tons, 20 times more than a typical car, and at 65 MPH takes a 10th of a mile to stop, almost twice as far as a car (Exhibit 6). Wind is a big problem for trucks – a 33MPH gust can displace a trailer by 10 feet when driving 45MPH – and drivers must be constantly attuned to the effects, sidelining their vehicle when conditions are unsafe (Exhibit 7). These mechanical beasts are also difficult to maneuver, particularly with smaller and nimbler cars and motorcycles sharing the roadway. A sharp turn can jackknife or roll a truck, damaging not just the truck but the valuable contents. An accident is much more likely to yield fatalities – large trucks are just 4% of US registered vehicles but were involved with 9% of all fatal road accidents (Exhibit 8). A minor accident can still cause hours of costly delay.
Exh 4: Long Haul Semi Truck Characteristics
Beyond the physical challenge of conducting an 18-wheeler, drivers have other responsibilities. The trucks must be inspected by the drivers – a broken tail light or a lost mud flap can result in a truck sidelined until a replacement part can be delivered and installed, along with a hefty fine. The driver must also present the truck for weighing and inspection at periodic road checks, with accompanying paperwork to be filed. At fueling stops, the truck is vulnerable to theft and vandalism, and drivers are required to be vigilant. Drivers must also monitor public safety channels and obey restrictions on truck travel during inclement weather and other conditions. Self-driving systems will have to account for all these responsibilities.
Exh 5: Average Lane Utilization by Vehicle Type
Exh 6: Stopping Distance by Vehicle Type and Speed
Exh 7: Trailer Displacement due to Strong Winds
Exh 8: Fatalities in Crashes Involving Large Trucks, 2010 – 2015
What’s an AI Gotta Do Around Here?
In our previous work on autonomous cars, we have described the technical challenges for self-driving systems, (http://www.ssrllc.com/publication/autonomous-cars-self-driving-ambition/) dividing the tasks into 3D Mapping and Autonomous Controls (Exhibit 9). In the first element, the system takes a detailed static digital map of the planned route and uses the fused input of electronic sensors (Lidar, cameras, radar, microphones, ultrasonic sensors, etc.) to generate a real time, 3D representation of the environment around the car, including estimations for the likely forward path for objects in motion. For trucks, this may be different than for cars in several ways (Exhibit 10). First, existing highway maps built for cars do not necessarily deliver details that could be important to trucks – overhead obstructions, precise lane widths, special restrictions for trucks, etc. Second, the size of trucks and their long stopping distance requires a much larger radius for sensors to map. Third, it is more important for trucking systems to assess the driving conditions (e.g. weather, uneven pavement, road damage, etc.) and their potential effect on the vehicle. For these reasons, 3D mapping solutions built for cars cannot be simply adapted for use on trucks. The systems must be built and trained specifically for the task (Exhibit 11).
Exh 9: Self-Driving Technology Elements
The autonomous control system is the brains of the self-driving vehicle. This is an AI model trained to react appropriately to the conditions presented by the 3D mapping system and make the decisions necessary to drive the truck. In cars, Google’s Waymo business is finally ready to offer fully driverless operations in Phoenix from its fleet of specially equipped minivans, after nearly a decade of development, more than 3.5M miles of autonomous road testing, tens of thousands of hours practicing particularly tricky situations on a private testing facility, and billions of miles in computer simulation. Of course, Waymo’s driving system could not be recalibrated for trucks, and nor could any of the other self-driving car solutions, which are in various stages of development. A system for trucks needs to start from scratch.
Exh 10: Differences in Conditions Encountered
Exh 11: Additional Considerations for Self-Driving Trucks
1, 2, 3, Testing, Testing ….
Autonomous driving systems for cars are quite far along in their development. Google began its initiative nearly a decade ago, and has transitioned through at least 5 generations of system development. It has logged more than 3.5M miles of autonomous driving on public roads through deliberately varied conditions (Exhibit 12). It has duplicated the most challenging circumstances at a private testing facility to run repeated live exercises. It has hired police and fire departments to run emergency drills for its self-driving cars to experience. It has built a sophisticated simulation program, and has run its autonomous control system through many billions of miles of subtle variations on those public driving experiences. It has meticulously mapped the communities where it is running its tests, noting curb heights and every pothole along the potential routes. After all of this, Waymo will begin offering fully autonomous taxi service to subscribers in the Phoenix area within short months (Exhibit 13).
Exh 12: Disruptors with1M+ Tested Miles
Exh 13: California Reported Self-Driving Activity
In contrast, self-driving trucks are in their infancy (Exhibit 14). The only fully autonomous test to date has been a 120-mile jaunt by a beer delivery truck powered by Otto driving in one lane over an uncrowded stretch of Nevada highway accompanied by an armada of support vehicles. Otherwise, developers are still collecting data from vehicles driven by humans, still far from driverless testing with ready drivers set to take over in case of failure. The relative success of Tesla’s autopilot feature has given some observers false confidence about how quickly 80,000-pound semis might be able or allowed to do the same thing. We also note that testing will be much more expensive with $175K big rigs than with $25K minivans. We believe that highway testing for full autonomy in trucks could take 5 years or more, with further delays in getting truck operators, regulators and insurers on board.
Exh 14: Autonomous Trucks – Top Players Progress Capture
A Little Help
Meanwhile, truck OEMs are not really working on full autonomy yet either. Daimler has begun testing autonomous trucks in Nevada, but has set 2025 as a target for Level 5 functionality. In the near term, it is focused on developing autonomous shuttle buses that can follow a simple fixed route. Peterbilt has cooperated with Uber/Otto and Waymo on self-driving tests and has begun work on advanced cruise control automated driving systems to support human drivers. Volvo is also working on driver assist technologies, such as enabling efficient platooning. Uber’s Otto is a bit more ambitious, looking to offer Level 4 autonomy – which can largely handle highway driving without an operator, but that would still require a driver as a failsafe – by 2019. Tesla has suggested that it will offer an electric semi-truck by the end of this year, with self-driving capabilities sometime thereafter. It has applied to California and Nevada for permits to begin testing. Waymo has begun testing with a single Peterbilt truck, and like others, seems to be focused on driver support systems that don’t yet replace the human behind the wheel. Starsky Robotics is the most advanced startup player, having completed its own 120-mile driverless delivery. It is also focused on driver assistance functions, eschewing LiDAR sensors in a bid to keep system costs down.
Of this group, Daimler is the only one to set a specific date to take the driver out of the cab, and it is projecting 8 years forward. Given the realities of development and testing, and the possible struggles to bring trucking operators, insurers and regulators on board for a more radical solution, the incremental approach is the likely path forward. This contrasts with passenger vehicles, where we continue to believe that on-demand autonomous TaaS subscriptions could be a reality in early markets before the end of the decade (and before the end of this year in Phoenix). These fleets will rely on highly detailed digital maps and be trained on the specific terrain, conditions, and driving habits of the locale, and offer service within strictly defined geofenced boundaries. Waymo is well ahead, but because each market will require specific mapping and testing before commercial service launch, would-be rivals like Uber, GM and Baidu could be close enough to get into the game further down the line (Exhibit 15).
Exh 15: SSR Timeline for Adoption of Vehicular Autonomy
The Airport Shuttle
While we see robo-taxis as the big near-term opportunity, autonomous driving technology will certainly show in other applications. Even with the reduced maneuverability and other challenges for larger vehicles, fixed route applications are likely viable with current technology – for example, an airport parking shuttle that follows a clearly defined, simple route or even a delivery van making regular deliveries to pre-determined locations with authorized recipients (Exhibit 16). These vehicles would rely on detailed mapping, stay restricted to a single lane, and stop at clearly defined and reserved locations. Indeed, Google’s first autonomous vehicle was a shuttle bus used to ferry employees around its Mountain View campus. The US Postal Service recently revealed its intention to use autonomous vehicles to deliver mail on fixed routes for rural customers, beginning in 2025. These applications will not be huge markets, but could add valuable data and experience to add to more ambitious trucking projects.
Exh 16: Fixed Route Shuttle Deployment
Unlike autonomous cars, where we expect disruptors like Google, Uber and Baidu to drive fleet-based on-demand TaaS ahead of the global auto industry, the development and adoption of self-driving tech for trucks will be incremental. Tech companies will participate as 3rd parties supplying components and software to ecosystems controlled by truck makers and their operator customers (Exhibit 17). The clearest opportunity is for the makers of LiDAR and other sensors, communications chips, and for the on-board computing systems needed to run the driver assistance solutions – Nvidia is an obvious beneficiary, as are Qualcomm, Freescale, Autoliv, Bosch, On, Xilinx, and perhaps, Intel and Texas Instruments. There are also many private firms working in the area, such as LiDAR pioneers Velodyne, Quanergy, Luminar.
Exh 17: Summary of Positioning and Initial Partnerships
Exh 18: Summary of Autonomous Vehicular Market Opportunity
Exh 19: SSR TMT AI/Self-Driving Heatmap