Is autonomous trucking finally on the horizon?
With established routes and over 3 million miles under their belt, Kodiak Robotics’ CTO thinks that 2024 is a pivotal year for getting autonomous trucks into middle mile logistics
There have been plenty of big promises made when it comes to self-driving Heavy Goods Vehicles (HGVs), but commercial success has proved more elusive. Several high-profile start-ups have gone from multi-billion-dollar valuations to leaving the autonomous truck sector altogether, such as Embark and TuSimple.
Andreas Wendel, Chief Technology Officer for driverless tech firm Kodiak Robotics, has no plans to be among those industry casualties. To avoid it, they are putting in the miles, over 3 million of them so far, with a big focus on repeatable, predictable routes in the middle mile – journeys where autonomous systems excel.
Machines in the middle mile?
When it comes to where self-driving HGVs are going to be most impactful and are closest to achieving wider commercial deployments, Wendel is clear: the middle mile.
“Computers are extremely good at doing, essentially, very repetitive tasks,” says Wendel, “things that can be done over and over again.” That makes the middle miles across the “highway portion - the long, often dull, dangerous kind of stretches of the road” - the main target for companies like Kodiak Robotics, as “less and less people want to do it,” creating “a really big driver shortage in these stretches”.
It is on defined routes where we have tested technology extensively
He notes that they have secured contracts from companies like Maersk, J.B. Hunt, Ikea and Ryder to haul goods on these kinds of routes, moving between key points or from production and distribution facilities to end points.
“It is on defined routes where we have tested technology extensively,” he explains, “for instance, Dallas to Houston, or we also drive Dallas to Atlanta every day.” These are largely scheduled, repeat routes, and therefore often less profitable for independent truckers, but also where self-driving systems can perform best.
“In most of these instances there's actually locations very close to the highway that we go to,” Wendel says, such as rest stops, or in the case of Ryder, they head to “third-party locations that are already built for trucks. We go there and then a human driver picks up the load and delivers it through the last mile to a distribution centre.”
Currently, these routes require a person on constant standby within the vehicle even as it drives itself. Regulation for self-driving trucks, where it does exist, covers autonomous driving up to what is known as ‘Level 4’, which allows the vehicle to fully self-drive, but legally requires a human presence. This limits the scope for potential cost savings, but Wendel feels the application in these segments is clear.
“Undeniably there is a cost benefit that [customers] hope to get at in the long run,” that isn’t fully realised today due to these limitations admits Wendel, however “even in the short run, it's actually extremely hard to find drivers.”
Current shortfalls of truckers in places like the US and EU run in the hundreds of thousands, and could head well into the millions in China, with the potential for these acute labour shortages to skyrocket over the next half decade if left unaddressed.
Do you want someone who is safe and reliable and can deliver it today?
This dynamic is pushing more companies to take autonomous HGVs seriously.
“Do you want someone who is safe and reliable and can deliver it today?” Wendel asks. That can be “a very, very hard combination to find,” in Kodiak’s target segments at this time, let alone if driver deficits worsen over the coming half decade.
How does a truck see the world?
The challenges of developing and deploying systems that can guide a 30-tonne big rig safely are substantive. This is why the recent leaps in the capacity of Artificial Intelligence (AI) algorithms to perceive, and predict, what is happening both on, and around, the road so important.
Now, you just throw more and more data on it, and you use the machine’s compute power to figure out those patterns on their own, and that's the really big unlock that we have seen in recent years
Kodiak currently uses a set of sensors that it can attach to commercial vehicles, including lidar and cameras, but the really clever part comes in how those systems combine to create a picture of what is happening, and just as importantly, what might develop on and around the road in the next few seconds.
“When I started working on computer vision,” Wendel notes, “a lot of the … salient points and indicators you had to handcraft,” which meant a painstaking task of examining the data and building the mathematical models manually. “Now, you just throw more and more data on it, and you use the machine’s compute power to figure out those patterns on their own, and that's the really big unlock that we have seen in recent years.”
That leap has allowed real time sensor fusion and an increase in testing and feedback loops that are powering a much faster rate of progress than was possible a decade ago.
Kodiak’s approach rests on two main systems, the “perception system that looks at the environment and figures out what is going on” and a “a motion planner” he explains.
Given all of these observations around me, how should I drive and what is a safe driving behaviour?
The former “uses a lot of discriminative AI, figuring out what do things look like? Is this a vehicle? Is it a human? What classes of vehicles are there? What speed are they going at? It really analyses the environment.”
Then, the motion planner “tries to figure out, given all of these observations around me, how should I drive and what is a safe driving behaviour?
“Based on what's going on, you plan a path … and then you need to perceive the reaction of the world to it and you close the loop,” he says by “working on analysing all the data that we get from these to say how does it compare to how a human is driving? Then any changes we rate as to would that have been better or worse? Would it have been safety critical or not? Then we go and make those changes in our system.”
Injecting unpredictable scenarios for predictable driving
However, roads are unpredictable, even over those long, monotonous middle miles Kodiak is focused on, and scenarios regularly emerge that the AI systems have never seen and aren’t in the databases.
“You see people cutting each other off and road rage out there,” for example Wendel says. “There is actually a lot of very special cases that you have never seen,” but “you need to be able to handle all of this and get it into your datasets, into your simulations, into your evaluations.”
What if you have a person on a highway that maybe you don't have a lot of samples of?
It is these “edge cases” where Wendel says they have faced some of their biggest challenges, but also “where we have made most of our advances in the in the most recent years.”
For these edge cases, simulations here have been a key means to drive forward the system’s ability to react, allowing Kodiak to “actually inject some of these behaviours,” but do so in software, rather than real life.
They can then ask, “what if you have a person on a highway that maybe you don't have a lot of samples of?” and run and re-run the system’s reaction, fine tuning it each time. “You create a new data point … that you have never seen,” in these scenarios, but one that can still “show we can actually be safe in those challenging and very rare situations.”
Will trucking turn towards self driving?
So, will the sector finally look to embrace self-driving solutions?
Well, as we covered in this previous story, the lack of qualified HGV drivers is already pretty severe. What’s more, reversing this trend will require action and investment now to ensure a pipeline in the years to come, and that doesn’t seem to be emerging. Regulations introduced by the EU in 2020 to ostensibly improve truck driver working conditions do not seem to have had an appreciable impact.
There is a job shortage for this. We have customers who are paying us for this today
Then, when it comes to boosting trucker pay packets, the focus within supply chains has very much shifted back towards keeping in control of costs rather than paying out. Many C-suite executives have returned to seeing supply chains as a cost centre and executives in a recent report said that operational costs and workforce management were key challenges for them that required investment.
However, that investment in both these pieces of research was more directed to automation than pay rises. The former survey, from EY, found that 48% planned to invest in driverless vehicles including forklifts and trucks.
“There is a job shortage for this. We have customers who are paying us for this today,” states Wendel, “so now it's really on us to show that we can safely take the driver out.”