Waymo Engineers Answer Questions About Autonomous Driving

Three Waymo engineers had been online for an AMA on reddit and answered questions from the forum participants. The engineers were Satish Jeyachandran (Head of Hardware), Mizuki McGrath (Engineering Director), and Nathaniel Fairfield (Distinguished Software Engineer).

Here are some of the new insights that I got:

Hardware Costs

Right the first question was about the hardware costs. The Waymo engineers didn’t mention dollar amounts but said that the costs are 50% lower than in the previous generation.

Note: even that Waymo has mentioned cost reductions with their own sensors of 90%, we’d probably can estimate the entire amount with four short range Honeycom LiDARs and a long range LiDAR on the roof, as well as cameras, radars and multiple AI chips to be in tthe range far North of 50,000 dollars.

Route Selection

When asked how the vehicle selects a route to the destination, Nathaniel listed the driving time of the various routes, the traffic, as well as construction sites and other delays. Seemingly, however, no consideration of particularly scenic routes.

Driverless

What exactly does Waymo mean by the term “driverless”? Nathaniel said that there is really no driver in the car anymore. The vehicles are connected to an operations center, where a fleet team can monitor and monitor the vehicles and send them specific information, but the team cannot brake or steer from a distance. This control is carried out by the vehicle alone.

Behavior in Catastrophes

The vehicles can detect water, smoke, dust etc. and slow down or even stop. The operations center also monitors local conditions at all times and can call all vehicles back to the depot, order them to stop safely at the roadside, or divert the vehicles to another route. There is no emergency stop, but it is possible to instruct the vehicles to stop at the side of the road as quickly as possible.

Key Figures

It’s a nuanced challenge, which key figures make sense. You can easily fool yourself that this or that method or statistic is the right one. That’s why Waymo uses a combination of methods to get an overall picture of driving safety. And these are constantly adapted and changed.

• Millions of kilometers driven on public roads and assessment of disengagements;
• Simulation of billions of miles, where an extensive set of challenging scenarios is tested, including those they have never seen before, but which could happen;
• Tests on closed test tracks so that rare scenarios can be simulated sufficiently often to obtain meaningful data;
• Rigorous design principles (redundant actuators, backup systems), security-based design, and extensive validation of system reactions;
• More details are in Waymo’s safety report;

Redundances

Waymo has built in the following redundancies:

• Power
• Braking/Steering
• Communication/Connectivity
• Sensors/Computers
• Onboard Software Stack

Machine Learning / Deep Learning

Waymo originally started with deep learning to recognize and classify objects, and now uses DL to predict actions of other road users, planning, mapping, and simulation. They work with other groups within Google, including Brain and Deep Mind.

Current Challenges

From a hardware perspective, the current challenges are that the sensors work optimally in all kinds of weather conditions. They have tested them in heavy rain in Florida, snow in Michigan, and dust storms in Arizona. They are currently working on the fifth version of the hardware equipment, which is currently undergoing final work.

Pedestrians

When asked how the vehicles react to unpredictable actions of pedestrians, the team said the vehicle behaved very carefully around them. It turned out that actions of the vehicle speak a clear language, especially if eye contact between car and person is not possible.

Processors

Waymo uses today CPUs, GPUs, accelerators and IO processors.

Use of External Sensors

Currently, Waymo has not encountered any scenarios in which sensors mounted on houses (to look around the corner), for example, have to be used by them.

Prediction Accuracy

To improve the accuracy of prediction, Waymo uses a large amount of data from the real world to measure prediction performance, on the one hand, and predicts several possibilities, on the other. They estimate the probability of each action, which is then considered by the planning module to create a safe plan.

When?

Waymo’s current primary focus is on robotic taxis and long-distance truck hauling with commercial B2B deliveries. For private vehicles, they are currently working with car manufacturers to install the Waymo Driver. Here they will make their own announcements with a cool business model.

Maps

The vehicles rely on maps, but the system is designed to navigate safely when roads have changed (e.g. construction sites). Road maps are useful to know where the vehicle is and predict what will come, such as stop signs around the bend that you cannot yet see visually.

Waymo creates their own high resolution 3D maps. The vehicles benefit from dynamically exchanging new information with the fleet, such as whether there is a construction site, or some new or temporary objects.

New Sensors

In the upcoming fifth version of the sensor suite, a new type of sensor is to be installed. Which one it is was not announced.

Swerving at Frontal Collisions

The vehicle has the freedom to decide whether it is safe to avoid an imminent frontal collision and swerve or not. It can do this in any case, and has been extensively tested on the test track.

Dark Objects

The LiDARs can also detect very dark objects, such as car tires in the dark on the freeway ahead of time and react safely. Dark cars also have sufficient reflective elements to detect them with LiDARs. And with radar, the color doesn’t matter, they always detect just as well.

Cameras

The cameras were developed by Waymo and are better than comparable cameras on the market.

Are Waymo Vehicle Communicating?

Yes, the Waymos share information with each other.

This article was also published in German.