Waymo has offered its first detailed look at the computing system inside its driverless vehicles, explaining what happens when the electronic driver encounters a failure. The company describes a machine built to make driving decisions in milliseconds without waiting for the cloud or a remote human operator, and it has layered backups into the system because there is no person in the front seat who can grab the wheel.
The newest compute system works somewhat like two independent engines. During normal driving, both computing systems can run full workloads in parallel. If one develops a fault, Waymo says the other can take over, giving the vehicle another electronic driver already onboard rather than forcing it to wait for someone miles away to intervene. The company's broader safety documentation adds that a secondary onboard computer can bring a vehicle to a safe stop if it detects a failure in the primary system.
Beyond computing, Waymo says the vehicles include backup braking and redundant steering hardware. Separate power sources support critical driving systems, so a single power problem does not automatically knock out everything the vehicle needs to operate. The priority in a serious failure is to keep the vehicle under control and reach a safe state, which could mean stopping rather than finishing the trip.
A Waymo takes in a huge amount of information as it moves. Its cameras feed images into the onboard system, lidar uses laser pulses to create a detailed picture of the surroundings, and radar helps measure how far away objects are and how quickly they are moving. The latest sixth-generation sensor system uses 13 cameras along with lidar and radar sensors. The computer then has to make sense of all that information fast enough to steer, brake, or react to something entering the road.
Waymo says those driving decisions happen entirely onboard. That point matters for anyone who has wondered whether a robotaxi needs to send everything to a giant data center before making a turn. It does not. The onboard computer handles the immediate driving workload itself. According to Waymo, it has increased its raw computing power about 20 times over the past eight years, which helps shrink the time between seeing something and reacting to it.
Waymo also revealed a new custom 5-nanometer ASIC, or application-specific integrated circuit. The chip handles raw information flowing in from the vehicle's sensors, processing camera, lidar, and radar data before that information reaches other parts of the driving system. Waymo says the ASIC delivers more than 1,000 TOPS of machine-learning performance, meaning trillions of operations per second. The practical idea is that the car can recognize what it sees and respond with as little delay as possible, including processing footage from all 13 high-resolution cameras at the same time. Waymo says its technology also improves what the cameras can detect in low light.
Computing redundancy only covers part of the problem. If the backup computer works perfectly but the vehicle suddenly loses its normal braking hardware, Waymo says a secondary braking system can bring the vehicle to a safe stop. Steering also has redundant hardware, with separate controllers and power supplies supporting the secondary steering system.
Waymo has been building these layers into its autonomous vehicles for years and has previously described testing scenarios involving hardware faults, power problems, and software failures. Its vehicles continually check their own systems while driving. That approach makes sense when the human backup has been removed from the equation.
One of the biggest misconceptions about driverless robotaxis is that an employee sits in an office with a steering wheel and pedals waiting to take control. Waymo does use what it calls Remote Assistance, but the company says those employees provide information rather than remotely driving the vehicle. If the Waymo Driver encounters an unusual situation, it can request additional context from a remote agent, then decide whether to use that information. Remote agents do not continuously watch every vehicle.