General Motors has noodled with the idea of automated vehicles since at least the mid-1950s, when the company produced a film with its Firebird II concept car that showed a family cruising down a highway autonomously. Now, more than six decades later, GM claims to be getting close to making its first highly automated vehicles a commercial reality.
Although it has shown glimpses of three generations of automated Chevrolet Bolt prototypes over the past 18 months, GM has remained remarkably mum about the technical details. On Monday, the automaker and its Cruise Automation subsidiary finally shared more about its current-generation program and allowed a small group of reporters to go for rides near its San Francisco headquarters.
As in Waymo’s automated Chrysler Pacificas, GM has equipped each seat with a tablet that provides a display of what the sensors are seeing, the paths the car is considering, and the currently chosen path. All of this is meant to help build trust by assuring riders that the car is “seeing” everything moving around the car.
The Cruise team built and has spent several months testing a smartphone app to request rides in the fashion of ride-hailing services Uber and Lyft. We were allowed to choose from several destinations to go for about a 20-minute ride in the Bolt EV, a vehicle positioned as a flagship for the company’s future-minded urban-mobility efforts.
Sitting in the back seat, I scanned back and forth between the screen in front of me and the surroundings to verify that everything I could see was being detected by the car’s sensors. Throughout the ride, the car took a fairly conservative approach, always erred on the side of safety, and always seemed to be choosing the optimal path. When it stopped behind a double-parked truck dropping off a load, it waited for another car to pass, and then peeked out to the left in order to “see” if anyone was coming from the other direction before proceeding to go around.
From an object-detection and path-planning perspective, the Cruise software seemed to work as well as any of the others we’ve sampled from Ford, Delphi, Waymo, and others. Where the Bolt EV still lags compared with others is in overall control refinement. Some of the steering, braking, and acceleration maneuvers felt rougher and more aggressive than one might want from, say, a Lyft driver. It felt like what I experienced when teaching my kids to drive as they were still trying to come to grips with how much steering or braking input is required to get the desired response.
In contrast to the automated control strategies, driving refinement is a comparatively easy problem to solve. GM president Dan Ammann, vice president of autonomous and electric-vehicle programs Doug Parks, and Cruise CEO Kyle Vogt emphasized repeatedly throughout the presentation and follow-up conversations that they have focused until now on foundational elements of the self-driving hierarchy of needs—in other words, get me where I want to go, and get me there safely.
From a brief experience in the car, that part seems to be working so far, but the refinement is something that will have to be addressed before these cars can start picking up passengers who aren’t GM or Cruise employees.
Steven Shladover, Director of California PATH, points out the magnetic sensors installed on the front of a Buick LeSabre at a demonstration of the automated highway system in July 1997 in San Diego. The sensors, which read magnets in the road, allowed the car to be controlled by an onboard computer.
A Lengthy History of Autonomous Efforts
Prior to the turn of the century, most companies were pursuing concepts that relied more on infrastructure to guide vehicles down the road. Much as automated guided vehicles, which have been in factories since the 1980s, follow wires buried in the floor, it was assumed that cars would follow cables embedded in the road. As late as 1997, GM conducted just such demonstrations on a stretch of highway near San Diego, California (shown above).
GM continued working on its automated driving systems internally, including the 2010 EN-V concepts and later prototypes based on the Opel Insignia. But it took the acquisition in early 2016 of a small San Francisco startup called Cruise Automation to kick things into overdrive. GM quickly churned out a fleet of 50 Chevrolet Bolt EVs equipped with sensing, actuation, and computing capabilities, enabling the Cruise team to apply their startup sensibilities to help bring GM’s work up to production readiness more rapidly. The original 40-person staff at Cruise has now grown tenfold.
From Research to Production Engineering
While the Cruise software engineers began working with that first batch of cars to understand the capabilities and limitations of the sensor suite, the hardware team at GM’s Warren, Michigan, tech center went to work on developing production prototypes. According to executive chief engineer Andrew Farah, the team followed two parallel paths. The first focused on putting together an even more capable, multilayered sensor suite consisting of more than 40 cameras, radars, and lidars. This led to the batch of 130 vehicles that were assembled at GM’s Orion, Michigan, assembly plant in May 2017.
“GM has a distinct advantage,” said GM’s Parks in San Francisco. “We can provide integration, a multilayered sensor suite, validation, and manufacturing scale.”
All of this is crucial to taking highly automated driving from the science project it has been for decades into a viable business capable of achieving the company’s oft stated goal of zero emissions, zero fatalities, and zero congestion. Any startup with $100,000 in seed funding can get a car, an Nvidia Drive PX computer, and a box of off-the-shelf sensors and get a halfway decent demo running on the streets near Sand Hill Road in Palo Alto, where all the big Silicon Valley venture capitalists are located.
“We need to get commercial deployment of autonomous vehicles at scale as soon as possible.”
– Dan Ammann, General Motors president
The problem is that such a system will have a hard time dealing with all the vagaries of real-world urban traffic in places like San Francisco, Manhattan, Singapore, and Shanghai. That’s where the experience and capabilities of a company that has been building vehicles as complex as cars for decades come into play.
With safety as the company’s top priority for its automated driving system, GM has gone all out on what it intends to use as its first production application. Five Velodyne Puck lidar sensors adorn the roof (photo above), augmented by 14 cameras. In addition to eight static long-range radar units, the Bolt EV has three articulated medium-range radars. These can be swiveled around to provide a more complete view of the surrounding areas as the car maneuvers in tight quarters, just as a human driver would look over his or her shoulder while backing up.
Finally, GM developed new ultra-short-range radar sensors, 10 of which are fitted around the car. GM executive chief engineer Farah explained that ultrasonic sensors, like those used for park assist or for Tesla’s Autopilot, give far less detail about objects that are detected. Essentially, ultrasonic sensors have a hemispherical field of view and only indicate if an object is within that field—but not where. These new radar sensors can detect where objects are vertically and horizontally and how far away they are.
Making the “Fail-Operational” Automated Vehicle
The second hardware engineering path at GM was focused on all the systems needed to make an automated vehicle fail-operational. With existing driver-assist systems, if a sensor, computer, or actuator fails, the human driver is there to take over. In a car that may not have a driver, major components require backup systems that can take over on the fly and bring the vehicle to a safe place. This hardware setup was combined with the sensor platform shown earlier and unveiled as the third-generation prototype just three months after generation two.
Farah said the team developed a new electric steering actuator containing two concentric motors. Normally, the motors work together to provide directional control, but in the event of a problem, either can take over and do the job. Similarly for the brakes, there is an electromechanical booster that provides primary brake application. The stability control system’s hydraulic actuator can also apply full braking force.
Cruise CEO Vogt declined to discuss most specifics about the computer platform being used in the Bolt EV, including what chips are being used. He did, however, acknowledge there are two primary identical computers running in parallel, along with a third with a different architecture that functions as a backup.
All of this hardware requires electrical power—and lots of it. Two independent power supplies ensure that all aspects of the car get their flow of electrons at all times. Farah acknowledged that the first-generation prototypes used about 3 to 4 kW of electrical power for the automated driving system and that the newer cars use less, but he declined to say how much more efficient they were.
“We need to get commercial deployment of autonomous vehicles at scale as soon as possible,” Ammann said. “Scale will be essential to achieve the benefits of this technology for society.” While he isn’t getting specific yet as to the timetable, he reiterated that it will be measured in quarters, not years.