When it comes to self-driving trucks, trucks that can tool down the highway without a human behind the wheel, the common perception is that this is a very complex technological problem. It is thought that solving the artificial intelligence problem, computer-based intelligence that can navigate roads as safely or more safely than a human, is a particularly hard nut to crack. Several pundits argue this technology is still probably ten years away. Stefan Seltz-Axmacher, the CEO of Starsky Robotics, begs to differ. “When it comes to safety engineering, in general, simpler is better.” Starsky is zigging while the rest of the industry zags.
He may be right. In a technology arms race that has seen hundreds of millions of dollars in investment, and the participation of much larger and better-known technology companies, Starsky Robotics has gotten out of the gate first. Yet Starsky Robotics, a San Francisco startup building driverless trucks, has only been funded to the tune of $21.7 million.
On June 16th, a Volvo semi-truck using Starsky technology did a 9.4-mile trip on the Florida Turnpike in Orlando. The unmanned truck with no human being on board successfully navigated through a highway rest area, merged onto the Turnpike, and changed lanes, while maintaining an average speed of 55 miles an hour.
Autonomous Trucking Requires Diverse Sensing Technologies
I’ve talked to other technology firms trying to solve the autonomous truck problem. They use LIDAR and look to combine it with cameras, and regular radar. Even ground penetrating radar is being touted as a redundant sensing technology. The idea is that sensor redundancy is necessary for safety and that different sensors perform well in different environmental situations.
Mr. Seltz-Axmacher believes that redundancy in sensors to promote safety is critical. But the more types of sensors involved, the more complicated the technology becomes. Starsky only relies on radar and cameras. When I expressed surprise that they were not using LIDAR, a sensor system that uses light from a laser in a similar way to the way in which radar uses electromagnetic waves, Mr. Seltz-Axmacher said “LIDAR! People have always questioned why we don’t use LIDAR. The answer is simple. Because we don’t need to.”
LIDAR, in combination with other sensors, is being used to help trucks navigate based on a technology known as simultaneous localization and mapping (SLAM). SLAM seeks to triangulate on objects beside the road to make a probabilistic decision on exactly where the vehicle is located. “This is theoretically a great idea,” Mr. Seltz-Axmacher says. “But LIDAR is a notoriously tricky technology.”
Starsky argues that SLAM is useful for navigating in the local environment (e.g. where I am relative to my lane) but not in the global environment (e.g. which exit I should take). High-precision GPS can replace SLAM for the global environment. However, other autonomous vehicle suppliers point out that GPS may suffer reduced performance or outages. SLAM can fill in the gaps in navigation in the case of these difficulties.
The toughest problem autonomous navigation systems seek to solve is navigating in urban environments, which is much more difficult than navigating on an interstate. In urban environments, pedestrians can cross streets unexpectedly, cyclists can engage in dangerous activities, potholes need to be avoided, and many other complexities exist.
In contrast, Kartik Tiwari, the Chief Technology Officer at Starsky pointed out in an article in The Drive, that “on interstates, fences prevent pedestrian and animal access. Vehicles enter and exit via smoothly curving on and off ramps. Road curvature is limited to what is easily navigable at 60 miles per hour, and pavement is smooth and well-maintained.”
Some autonomous vehicle companies believe that even trucks that travel on the interstate need to have the kind of robust autonomous navigation capabilities needed in urban areas. The reasoning is that stupid things do happen. Pedestrians do attempt to run across a highway, or cyclists may be riding on the berm, or animals do manage to get on the highway.
Starsky tackles these kinds of issues in a different way. In “non-normal” driving situations, the truck goes into a safety state called “minimal risk condition.” This state allows a vehicle to minimize its risk. Most often that will mean pulling over on side of road. The minimal risk condition can be triggered by the driving environment, or it can be caused if connectivity to the GPS system fails or if other sensors fail.
A Different Business Model for Autonomous Trucking
The question of what an autonomous vehicle’s technology is designed to do, also gets tied up with the business model for the autonomous trucking firm. Embark Trucks’ business model is on-ramp to off-ramp. A driver takes a truck from a distribution center onto the highway, parks it a truck stop on the Interstate, and then pushes a button and the truck travels to a truck stop on an interstate near the destination distribution center. There the truck is greeted by a driver that drives the truck the final few miles to the distribution center.
Mr. Seltz-Axmacher’s view is that adds too much complexity to logistics and adversely impacts the ROI of the solution. Starsky trucks will travel from distribution center to distribution center, on the highway as well as on the local roads to the warehouse.
How can they hope to do this without the kind of robust autonomous navigation being developed for off-highway usage? By using human drivers in control rooms connected to trucks over the cellular network. These remote human drivers will control the trucks from warehouse right up until the time the semi enters the highway and help safely navigate complex, context-heavy traffic environments.
Mr. Tiwari, the CTO, argues that “building autonomy to operate safely on highways is easier than building autonomy to operate safely in truckyards, or any complex, low-speed environment. Precisely the inverse is true for tele-operation of vehicles. The big limiting factor for tele-operation is the relationship between signal latency and speed. Signal latency refers to the gap between the moment the human tele-operator directs the vehicle to do something, and the moment that the truck actually does it. The higher the vehicle speed, the more problematic signal latency becomes for tele-operation.” In short, autonomy works well in the more tightly controlled interstate highway environment, and tele-operations works well in highly complex low speed environments.
Mr. Seltz-Axmacher believes their simpler is better approach will allow them to win the race to put this technology into everyday usage. Their goal is to have from 5-50 autonomous trucks next year. The goal is to ramp up production from that point forward. Mr. Seltz-Axmacher believes that the hard part will not be proving the technology. It will be scaling production and ramping up the new routes that are tested and validated as ready to use. “By 2025, there will be thousands of self-driving trucks on the roads.”