Why did Waymo's Demo only take 18 months, and why did the product take about 15 years?

This article is part of the PRO Members Communication, and follows "PRO Members of Machine Heart" at the end of the article to read more on the topic.

Autonomous driving was once thought to be rapidly becoming widespread, and now fully autonomous taxis are only operating in a handful of cities. Waymo's co-CEO Dmitri Dolgov recalled the team's research and development process from the early Demo to the massive operation of autonomous services. His team tested the ability of vehicles to complete autonomous driving tasks in just about 18 months, but it took about 15 years to build and expand autonomous services. To build a long-term reliable autonomous driving system, product research and enable the development of Demo's capabilities in an open environment and team-scale environment, continuous processing of hardware aging, transmission sensors, road changes and rare events, decided how Waymo would combine different sensor designs, verifiable model structures, and complete product evaluation systems through simulation and interaction.

Directory

01. Autonomous driving has already accomplished its mission, so why can't it be a direct product?

Autonomous driving has already accomplished its mission, so why can't it be a direct product?

Dmitri Dolgov, co-CEO of Waymo, recently delivered a speech at the Y Combinator Startup School, reviewing Waymo's R&D process from early capability verification, on-line driverless services to fleet scale expansion. He summarized the approximately 15-year product development process as a continuous improvement in long-term reliability, enabling systems to repeatedly complete tasks in open environments, different cities, and expanded fleet sizes, and to cope with component failures and low-frequency abnormalities.

1 Dmitri Dolgov is one of the founding members of the Google Autonomous Vehicle Project, formerly Waymo CTO, now responsible for corporate strategy and R&D and deployment of Waymo Driver; prior to joining the project, he was involved in autonomous driving research for the Toyota and Stanford University DARPA Urban Challenge teams.

2 By 2010, the vehicle had reached Dolgov's Capability Complete, with the main capabilities needed to complete the demonstration mission; about a decade later, Waymo began offering driverless service and would expand to its current size in about five years.

2.Dolgov first distinguished between Demo and product evaluation standards. Early Demo verification systems required the ability to complete unmanned routes under a variety of road conditions, which required long-term stable operation in open environments. Evaluation standards were incorporated into the design and verification of vehicle aging, sensor blindness, road changes and abnormal events after a single successful transition to continuous driverless driving.

The team, consisting of about 12 people, completed two goals in 18 months: a cumulative 100 thousand miles, and a complete uninterrupted run on 10 routes of about 100 miles in length. Tests covering day and night, normal roads, highways and construction routes verified the vehicle's basic ability to complete autonomous driving tasks.

3.Dolgov spoke of the fact that self-driving systems operate directly on the physical world, a characteristic that naturally widened the gap between demonstrations and products. He argued that there were four main gaps between physical AI and digital AI: error costs are higher and vehicle errors are irreversible, response time is shorter and decisions must be made in milliseconds, training data is more difficult to obtain, verification requirements are more stringent, and products must be fully verified before they are operated by humans.

1 Digital assistants can be reproduced after incorrect responses are given, but the wrong decisions of the vehicle can result in personal injury. For example, Dolgov says that a car can advance about 100 feet in a second at high speed, and the vehicle system needs to make perception, reasoning, and decision-making in milliseconds.

2 Digital AI can be used to generate user feedback from text, images, and products accumulated over the Internet, and the physical world lacks data that is comparable in size and contains action results.

4. Restrictions such as faulty costs, response speeds, and so on can only prove that the system has a basic task capability in an uninterrupted process. Waymo needs to transform this capability into deployable products, increase reliability from 90% to 99%, 99.9%, and continue to reduce system failures through redundant design, failure reduction and stricter verification.

1 According to Dolgov, every time reliability increases by a "9", the difficulty of achieving it increases by about a quantitative level. The rare occurrence of bicycles running millions of miles per week may occur even after a team runs millions of miles per week.

5. Reliability begins after the evaluation standards are shifted to long-term unmanned adoption and the performance limits and fault tolerance of the technical routes are predetermined. Waymo needs to judge what information different types of sensors can provide, what capacity limits models can achieve, and how the system will continue to operate after the component fails.

How does reliability change Waymo's technological path?

1. In order for the driverless system to continue operating in an open environment and on a fleet scale, Waymo assesses the performance of the prototype phase at the same time as the reliability thresholds that can be reached after a continuous increase in data, computing and engineering inputs. Dolgov notes that many systems progress rapidly in early testing, but the platform thresholds cannot support long-term unmanned operations if they are lower than product deployment requirements.