eyes on the road How GMs Super Cruise went from limo driving to lane changes and towing A self-learning algorithm runs on each car to understand the driving conditions. Jonathan M. Gitlin Jan 13, 2025 2:48 pm | 29 Credit: General Motors Credit: General Motors Story textSizeSmallStandardLargeWidth *StandardWideLinksStandardOrange* Subscribers only Learn moreWhen we first tested Super Cruisein 2018, the partially automated driver's assist impressed us enough that we wanted to see it rolled out across as much of General Motors' lineup as possible. Seven years later, our attitude toward such driver assists is a little more sober. Drivers are often more confident about such systems than they ought to be, and that's when they even care about such features in the first place.That said, Super Cruise remains one of the better implementations of what the industry has inelegantly labeled "level 2+" driver assists: in plain English, a system that lets the driver go hands-free for long stretches, as long as they're paying attention to the road. Which, in Super Cruise's case, is achieved via an infrared camera that faces the driver and follows their gaze, even if they're wearing sunglasses.Better yet, it's also tightly geofenced, as it's only meant to be used on restricted access, divided-lane highways.Even with this rather rigid "operational device domain," Super Cruise still has to be able to cope with a wide range of driving conditionsfrom the depths of a northern winter to the heat of a southwestern summeras well as different vehicle configurations."Since day one, we designed the Super Cruise steering control with multiple objectivesmainly to be precise, comfortable, and confidence-inspiring," said Reza Zarringhalam, a staff software developer at General Motors. "We... wanted to make sure that the drivers feel comfortable when the system is doing the driving task for them," he said.That means the lane centering can't have the driver thinking they're too close to, say, a semi the next lane over. The automated lane changesnow initiated either by the driver using a turn signal or if the car itself determines there's less trafficneed to be smooth. And importantly, it should be easy for the human in the driver's seat to take back control of the steering without having to fight or battle a machine.The Unified Lateral ControllerThe algorithm that handles all of that is called the Unified Lateral Controller. "So it's a single software stack, but it is also modular to adapt with different vehicle configurations, with different driving scenarios, different maneuvers," Zarringhalam said."Let's imagine that you're driving a Super Cruise vehicle, and you indicate to the left, or the system automatically decides to make a lane change to the left, and then, for whatever reason, the driver decides that they want to go back, mid-maneuver; they want to go back to the original lane. So you can just indicate to the opposite side, in this case, the right-hand side. Under the hood, in this scenario, everything is jumping. Our target trajectory is jumping from a left-lane maneuver to a right turn. The turn can be very sharp. There could be other objects that narrow the envelope of operation that you're allowed to function in," Zarringhalam said.Again, that behavior has to be consistent and predictable, whether it's below freezing or in the middle of a heatwave, and things like tire wear must also be taken into account. Or, say, the presence of a trailer, which could be anything from a bike rack with wheels to a three-axle trailer."As soon as we detect that the trailer is attached, we run several real-time algorithmstrailer inertial parameters, trailer math, trailer configuration, even how many axles we have, and the control adapts itself to execute lane turning and keep both the vehicle and the trailer at the center of the road," Zarringhalam said.That's done automatically without the driver having to input the information (obviating the problem of someone entering the wrong details), "and if you change the loading or the trailer configuration, even mid-driveif you pull over, load more weight and continue driving on the same road with Super Cruise activethese learnings happen in a matter of seconds," Zarringhalam said.Light machine learning"We created this idea of a self-learning vehicle control algorithm. It directly and indirectly observes the main parameters that we require for control," said Ali Shahriari, staff software developer at GM. "When you try to do a curve-tracking, the amount of steering you need to apply is different when a trailer is attached or when a trailer is not attached. So what the system is doing is trying to do this determination, quantify the impact, back-calculate the impact that controls will need to understand, and provide that information to the controls," Shahriari said.The self-learning algorithm is also how Super Cruise deals with changing climate conditions. "Giving an extreme example, let's say that you're on a low-friction surface. When you turn, the vehicle curvature is different. This is instantaneously learned and sent to the control, so [the car] knows that the response is different, so now [it] needs to make the right decision, so [the system] always provides optimal control," Shahriari said. Late last year, GM drove a convoy of 20 Super Cruise-enabled vehicles across the Bay Bridge in California. Credit: General Motors The calculations are all done locally, running on each car's onboard ECUsShahriari describes the self-learning algorithm as "light machine learning," as it's meant to run in real time on the car. "So when it's light, it's multiple different non-linear, linear regressions, adaptive filters, that explicitly monitor all of the learning conditions that are appropriate to learn on each car," Shahriari said.An added benefitfor the automaker, at leastis that the self-learning algorithm has made it a lot faster for GM to deploy Super Cruise across more and more models, reducing development time by as much as two-thirds. "It reduces complexity. It means software also is less calibration-intensive, because now it's all self-learning, and it's robust," Shahriari said.Jonathan M. GitlinAutomotive EditorJonathan M. GitlinAutomotive Editor Jonathan is the Automotive Editor at Ars Technica. He has a BSc and PhD in Pharmacology. In 2014 he decided to indulge his lifelong passion for the car by leaving the National Human Genome Research Institute and launching Ars Technica's automotive coverage. He lives in Washington, DC. 29 Comments