Two Waymo driverless taxis stop before passing one another on a San Francisco street on Feb. 15, 2023.Terry Chea/The Associated Press
In a split-second manoeuvre, a driverless vehicle swerves around a pedestrian who fell off their scooter and into its path. An Instagram post from December shows the vehicle narrowly avoiding the person – a real-world demonstration of how much self-driving technology has improved in a short time.
The vehicle, equipped with the Waymo Driver self-driving technology including cameras, recorded the incident in Austin, Texas, showing the vehicle’s hyperawareness of its surroundings. The instant reaction time was governed entirely by sensors, data and artificial intelligence.
David Margines, director of product management at California-based Waymo, a subsidiary of Alphabet Inc., says the technology was designed to improve safety for all road users to avoid accidents like what may have happened in this incident if the car didn’t swerve.
“In this instance, you can see the vehicle had slowed down and was planning to give them extra space while passing well before they began to stumble,” Margines says. “We’re proud that our behaviour prediction and fast reaction time helped that person get home safely that night.”
Driverless vehicle technology has received a major boost recently owing to advancements in AI and machine learning. Waymo claims to be one of the first companies to implement machine learning into its software stack, but AI’s role has grown massively in recent years as the field progresses.
Steven Waslander, a professor at the University of Toronto’s Institute for Aerospace Studies, who has been working on self-driving cars since 2014, says AI always been in the mix but only recently has it become more reliable thanks to new research and investment.
“Every year, the latest and greatest networks, modifications and revisions are being added from the research community into the vehicles,” he says. “There’s this constant flow from academia into companies like Waymo.”
Some of the newest developments include using spoken language and ChatGPT to help solve unique and difficult situations driverless vehicles may have yet to encounter, such as a blocked or obstructed roadway. Passengers will be able to describe the scene to ChapGPT and ask it for suggestions or give the vehicle instructions to clear an obstacle.
Predicting movements
Waymo’s ride-hailing service, Waymo One, uses the Jaguar I-Pace electric SUV outfitted with its self-driving hardware and software. The service started in Phoenix in 2020 and now runs in San Francisco, Los Angeles and Austin with plans to expand to Atlanta and Miami. Using a combination of LiDAR (Light Detection and Ranging), radar and cameras, the Waymo driver, as the company refers to its tech, can see in all directions at once and with incredibly rich detail.
Each city that it operates in gets studied and mapped, which can take more than a year, according to Waslander. The company builds these maps, which it says have much more detail than Google Maps. For example, he says the maps are higher resolution and include details such as how tall the trees are and where all the stop signs, traffic lights, parking spots and addresses are. The maps are loaded into the “driver” and used in combination with the other sensors to provide a real-time view.
“Where AI really comes in is detecting, tracking and predicting everything that’s moving in the scene,” Waslander says, which in a congested city such as San Francisco can include vehicles, pedestrians, cyclists, fire trucks, ambulances and everything in between.
It’s referred to as a “deep-learning” network or “neural” network, Waslander says, which he describes as techniques to learn how to look at camera, LiDAR and radar data to find and track moving objects and track them over time. It can also be used to mimic human driving behaviour by studying millions of kilometres of driving data.
“From there you make predictions and plan how you’re going to interact with those predictions,” Waslander says.
Waymo publishes extensive data on its research and findings, especially in the areas that demonstrate the safety of its system. Data from more than 56.7 million miles through May 2025 showed that compared with a human driver, the Waymo Driver had 92-per-cent fewer crashes with injuries to pedestrians and 82-per-cent fewer crashes with injuries to motorcyclists and cyclists.
Its findings also found 96-per-cent fewer crashes involving injuries at intersections, which are the leading cause of severe road harm to humans, according to the NHTSA, and 85-per-cent fewer instances of crashes reported with serious injuries.
“They haven’t driven enough miles to say anything statistically significant about fatalities, but they haven’t had one [death they were responsible for] yet through millions of miles,” Waslander says.
Waymo has a full simulation environment that can run millions of kilometres at night in every city it operates that examines fleet-wide statistics, tracking failures, disengagements, near misses and dangerous situations. Before releasing revised code to the fleets, it also does driver tests where it puts the vehicles through pre-defined scenarios known to be difficult to it.
“It’s a bit like how airlines got safer over the years because they have a safety board that reviews every accident,” Waslander says.
Winter still a challenge
Winter weather is still a challenge for self-driving cars. While Waymo currently operates in warmer climates in cities such as Phoenix and San Francisco, it’s move north has been slower. It’s currently testing in New York and Michigan but has no plans to expand into Canada.
Waslander does spend time in Canada though, leading the Canadian Adverse Driving Conditions dataset (CADC) research program at the University of Waterloo, which researches self-driving in winter weather. The program created the Autonomoose, an autonomous vehicle based on a Lincoln MKZ that can be operated on public roads. It’s the first of its kind created by a Canadian university and it’s been designed to drive autonomously in snowy conditions and collect data.
Waslander says LiDAR data can be filtered to remove snow and precipitation and camera images don’t have to be modified to detect objects. He says it’s as though the vehicle can see through the snow.
Waslander says Waymo needs time to develop its technology to work in winter climates. “Our research has shown [winter driving] is not an insurmountable task.”