How Waymo is using AI for autonomous driving?

With the autonomous driving industry evolving further with every passing year, from Tesla and its exclusive driving aids to platforms like AutoX and Didi which already manage fleets of fully autonomous cabs, there has been a constant revolution in the sector. 

 

The steady technological updates in the areas of artificial intelligence, machine learning, and other sensors such as LIDAR and GPS have allowed manufacturers to enhance self-driving potential in cars. Of course, the autonomy levels differ for every player in the industry, but the goal of integrating advanced control systems in the vehicles, so that the sensory inputs can be interpreted for detecting signboards and steering clear of collisions, remains constant.  

 

What started off as the tech giant Google’s exploration of self-driving vehicles has now emerged as the platform - Waymo, which creates driverless vehicles that can safely deliver people from one point to another. The platform’s 360-degree perception technology allows it to identify obstacles like pedestrians, other vehicles, or any construction work from up to several hundred yards away. 

 

About Waymo

 

Vision - "make it easy and safe for people and objects to move around." 

 

It all kickstarted in 2009 when Google set up its self-driving car project, which later went on to become what we call Waymo.

 

Waymo dreams of a future in which there are minimal accidents triggered through distracted, careless, or impaired human drivers and also minimal traffic, a world where it is easier for the public to get around.  

Waymo has termed itself as a self-driving technology company rather than a self-driving car company. Having a fleet of thousands of self-driving cars, powered with sensors like lidar and enabling driver-free operations. 

 

The platform has partnered with many car manufacturers ranging from Jaguar Land Rover to Volvo for integrating its self-driving technology into its vehicles. The platform has also partnered with OEMs to expose their Driver to more places and people, across a variety of vehicle forms, be it passenger to commercial vehicles. 

 

Heading the company over the past five and half years, John Krafcik recently stepped down from the role of CEO in April 2021.  The CEO duties will now be essayed by the platform’s two current company executives — Tekedra Mawakana and Dmitri Dolgov. 

 

Waymo operates a ride-hailing service - Waymo One, that serves rides in Metro Phoenix, Arizona, each day. The service aims to help the riders in the area of Metro Phoenix in getting to where they wish to go, be it at a grocery store or for an outing spot.  In October 2020, the platform expanded the service to the public, and it is the only self-driving commercial service that functions in the absence of safety backup drivers in the vehicle.

 

The platform is presently testing its fleet of trucks in California, Arizona, New Mexico, and Texas, and has launched a pilot program for local delivery in the Phoenix, Arizona area.

 

In March 2020, Waymo Via, the trucking division of Waymo, was launched. This division is mainly focused on the transportation of commercial goods. The platform’s autonomous driving solution has been established for enhancing safety and boosting efficiency while the local delivery solution has been generated for revamping the experience.

 

In early October 2020, Waymo announced its plan of launching fully driverless vehicles to the public. This endeavor sprouted at a time when the COVID19 induced global pandemic necessitated limited person-to-person contact. 

 

Recommended blog - Impact of Electric Vehicles on Environment

 

The platform has developed a Simulation City system for testing its autonomous vehicles and to brace them for the real world. Waymo had initially been making use of CarCraft since 2017, but owing to some discrepancies in the programme, the platform opted for a second programme. 

 

Both these programmes are now adopted by the platform for testing their cars in Simulation City, in which the vehicles get tested and trained. The system also assists in validating the input data fed into the “Waymo driver” software. 

 

How does the Waymo Driver Work?

 

As pointed out by the platform’s website, 4 steps are followed by the Waymo driver as it operates. 

 

1. Mapping the territory

 

Prior to operating in an alien area, the Waymo Driver first maps the territory with a high rate of detailing noting everything from the stop signs to the curbs. Then rather than depending only on tricky external data like GPS, where the risk of losing signal strength remains, the platform adopts these detailed and intricate custom maps, which are equipped with real-time sensor data and are capable of determining the exact location of the road at any point of time.  

 

 

2. Overseeing everything simultaneously

 

Waymo Driver’s perception system takes in the complex data accumulated from its leading suite of sensors and comprehends everything around be it other vehicles to pedestrians or cyclists, using technologies such as machine learning. At the same time, it also pays heed to signals and signs such as stop signs or the switching colors of traffic lights. 

---

How the Waymo Driver works

---

3. Predicting things before they occur

 

The driving scenario involves a range of objects each having its own behaviors and motives. The Waymo Driver gathers the information in real-time, mixed with its experience of real-world driving for gauging what other road users might do. It gauges how a car operates as opposed to that of a bike, a pedestrian, or any other object and then predicts pathways that the remaining road users can take up, within seconds.

 

 

4. Planning for the safest outcome

 

The Waymo Driver gathers all the information, be it the detailed maps, the objects in the surroundings, and where they might go, and plans out the best route or approach which can be taken. The trajectory, lane, speed, and steering maneuvers required for safely completing the journey are all determined in a flash. 

 

How Waymo uses AI in its operations

 

Car manufacturers across the globe have been making use of AI in each aspect of their car manufacturing procedure. Be it by robots assembling the initial nuts and bolts of vehicles or through autonomous cars making use of machine learning and vision for safely maneuvering their way through the traffic, AI has become an inevitable technology for the automotive industry. 

 

Recommended blog - Manufacturing in Tesla

 

Waymo considerably relies upon its AI research team, Google Brain, for integrating AI and ML capacity into its self-driving vehicles.

 

“We used a new machine learning technique called Deep Learning with help from the Google Brain team to teach our cars about the nature of objects on the road and understand how they would react,” 

- Dmitri Dolgov, Co-CEO, Waymo

 

The platform’s self-driving engineers operated in coordination with the Google Brain team for applying deep nets to the pedestrian detection system of the car. 

 

“While perception is the most mature area for deep learning, we also use deep nets for everything from prediction to planning to mapping and simulation.”

 

Machine Learning allows the cars to navigate complex and delicate situations such as operating through construction zones, relenting for emergency vehicles, and offering room to cars for parallel parking. 

 

 

1. Adopting the Tensor Flow ecosystem

 

Waymo makes use of the TensorFlow ecosystem and the data centers of Google such as TPUs for training its neural networks. TPUs (tensor processing units) allows the platform to train their nets up to 15 times with higher efficiency. Waymo also tests its ML models in simulation. Their rigorous testing and training cycle allows the platform to enhance its ML models and swiftly make use of the latest nets on its self-driving cars. 

 

 

2. Operate in challenging weather conditions

 

We are already aware of how strenuous the task of driving in heavy rain and snow is, both in the case of self and manual driving, owing to the lack of visibility. Waymo has trained its cars to operate in challenging weather conditions. 

 

Since snowflakes and raindrops can generate a great deal of noise in sensor data for self-driving vehicles, machine learning plays a part in filtering out the noise and also in properly detecting pedestrians, vehicles, and other objects. 

 

 

3. Simulate autonomous vehicle camera data - SurfelGAN

 

Recently the platform has announced its initiative of leveraging AI for generating camera images to simulate through sensor data gathered through its self-driving vehicles. 

 

In a recent paper that was co-authored by Waymo researchers including Research Head Dragomir Anguelov, the technique - SurfelGAN, has been described and elaborated in detail. SurfelGAN makes use of texture-mapped surface elements for reconstructing scenes and camera viewpoints to handle positions and orientations. The technique preserves sensor information while also saving considerable computational efficiency.

 

Waymo and similar platforms make use of simulation environments with the purpose of training, testing, and validating their systems prior to them being deployed for real-world cars. Waymo’s CarCraft is more computationally demanding since the programme attempts to model materials with a high degree of accuracy in order to ensure that the sensors such as lidars and radars operate authentically.

 

Recommended blog - Renewable Energy

 

Through SurfelGAN, Waymo has put forth an easier and more data-driven path to simulate sensor data. Through feeds gained from real-world lidar sensors and cameras, the AI generates and preserves exclusive data regarding the semantics, 3D geometry, and object appearances on the scene. Through the reconstruction, the simulated scene is rendered by SurfelGAN through varying distances as well as viewing angles.

 

Alongside AI, Waymo makes use of a range of technologies to enhance the exclusive experience offered by their vehicles. It remains to be seen how the autonomous vehicle sector will shape up in the future.