This interview covers Waymo's tech and business progress. The key message: the real breakthrough in self-driving isn't about "trolley problems" but raising the system's overall accuracy to avoid dangerous situations entirely. Dmitri Dolgov, Waymo's CTO, is confident in its multi-sensor approach (lidar + cameras + radar), arguing that ditching any sensor is "handicapping yourself." As for holdings: Waymo raised $3.2B and runs paid driverless rides in 50 sq miles of Phoenix, with demand exceeding supply; the Jaguar I-PACE is its fifth-gen hardware testbed; Tesla is implicitly criticized—Dolgov refutes Musk's view that lidar is a "crutch," insisting sensor fusion is safer.
This dialogue revolves around Waymo CTO Dmitri Dolgov's discussion on the future of autonomous vehicles. The core view is that Waymo, which began as Google's self-driving car project in 2009 and became an independent entity in 2016, currently holds a leading position in the full self-driving space,
The report, reviewed and reorganized by a third-party independent analyst, is based on the transcript of a podcast featuring Dmitri Dolgov of Waymo. The following is the analysis of the interview.
Guest Identity and Background: Dmitri Dolgov is the Chief Technology Officer of Waymo. He joined the Google self-driving car project at its inception in 2009 and has witnessed its entire evolution from a research project to an independent company.
Episode Main Theme: The interview delves into Waymo's complete technological evolution and commercialization path—from the DARPA Challenges and Google's internal project to independent operation and the realization of large-scale fully autonomous commercial services.
The Most Weighty Judgment in the Entire Episode: Dmitri Dolgov believes that the true breakthrough in autonomous driving technology does not lie in philosophical debates over 'crash choices' (such as the trolley problem), but rather in continuously improving the system's core capabilities (hardware and software) to shift the precision-recall curve upward as a whole, fundamentally avoiding the need to make such difficult decisions.
Dmitri Dolgov believes that Waymo's origins can be traced back to the DARPA Urban Challenge, an experience that not only validated the technical feasibility but also shaped the company's technical DNA.
Dolgov joined the Stanford University DARPA Urban Challenge team in 2006, responsible for motion planning algorithms. He recalled the 'magic moment' of running his own code on a real vehicle for the first time, and the anecdote of the team missing the championship in the 2007 competition due to a 'victory lap' bug. He believes that the success of these early challenges attracted the attention of Larry Page and Sergey Brin, thus giving birth to the Google self-driving car project in 2009.
In the early stages of the project, the team consisted of only about 12 people, and the goal was not to immediately build a product but to maximize learning efficiency. They set two key milestones:
1. Completing 100,000 miles of autonomous driving, an unprecedented scale at the time.
2. Achieving zero human intervention on 10 carefully selected, highly challenging routes, each 100 miles long, covering complex scenarios such as highways, mountain roads (e.g., Lake Tahoe), and dense urban areas (e.g., downtown San Francisco).
The team completed both milestones in less than two years (about 16 months). Dolgov pointed out that the core takeaway from this phase was 'the conviction that this is doable' (we believe that it's doable), although their understanding of the problem's complexity was still shallow at the time. In 2013, the team made a key strategic pivot, abandoning the plan to develop a Level 3 highway driver-assistance system and instead fully committing to building a 'virtual driver' capable of completely replacing a human driver in any situation. He believes this was an absolutely correct decision, because the technical paths for building a driver-assistance system and a fully autonomous system are fundamentally different.
Dmitri Dolgov emphasizes that Waymo's core strategy is the relentless pursuit of a fully autonomous state with 'no one in the driver's seat,' and this series of milestones constitutes the company's core competitive advantage.
Dolgov outlined the key milestones:
Dolgov believes that when passengers see the steering wheel turning on its own while the driver's seat is empty, the experience is 'transformative,' truly unlocking the potential of this technology. As of the time of the interview (December 2020), Waymo was offering fully autonomous paid ride-hailing services to the public in an area of approximately 50 square miles in Phoenix, with demand far exceeding supply. User feedback has been extremely positive, with the vast majority of ratings being five stars.
Dmitri Dolgov believes that Waymo's technical advantages stem from deep vertical integration of hardware (especially sensors) and the extensive application of machine learning across all components of the system. This stands in stark contrast to the 'pure vision-only' approach.
Dolgov elaborated on their technical system:
1. Hardware Self-Sufficiency: Waymo's lidar is almost entirely designed, engineered, and manufactured in-house, with core technology unique to Waymo. Radar and cameras also heavily utilize custom designs. The fifth-generation hardware system was designed with feasibility of mass production, unit economics, and higher reliability and redundancy in mind. Vehicle configuration includes 29 cameras, 5 lidar units, and 6 radar units.
2. V2X and Data Loop: Vehicles do not rely entirely on real-time communication; core decisions (such as safety-critical and latency-sensitive operations) are performed independently by the onboard system. However, vehicles share more a priori information (such as map updates, accidents, construction zones) and upload data in real time, enabling the entire fleet to 'collectively learn.' Additionally, the company has built a massive offline computing platform for training models, evaluating the system, and mining 'long-tail' edge cases.
3. Global Application of Machine Learning: Dolgov explicitly stated that machine learning is not only used for perception (e.g., object detection, classification) but is pervasive throughout the entire system, including prediction (behavior of pedestrians, cyclists, other vehicles), decision planning, and generative modeling in simulators. He specifically mentioned that breakthroughs such as the Transformer architecture in natural language processing are being successfully applied to behavior prediction and planning, because driving behavior itself is highly sequential and structured.
4. Combining 'Intuition' and 'Rules': Using the example of traffic lights, he pointed out that 'stop at red' is a clear, engineered semantic rule that can be directly injected into the system. But judging whether another vehicle will run a red light is a machine learning problem requiring extensive data training. Therefore, the optimal approach is to combine model-based systems with data-driven learning.
In response to Musk's view that 'lidar is a crutch,' Dolgov offered a strong rebuttal. He believes that lidar, cameras, and radar have different physical properties and complement each other, together building a safer and more powerful system. 'Why the heck would you handicap yourself and not use one or more of those sensing modalities?' He devoted considerable space to refuting two common criticisms: cost and aesthetics. On cost, with technological progress and mass production, lidar costs have dropped significantly. On aesthetics, he believes it can be made visually appealing through design.
Dmitri Dolgov believes that the Phoenix project is Waymo's 'platform' and 'stress test' for scaling expansion. Its core task is to complete a 'complete loop' from technical validation to product and commercial operations, thereby distilling a replicable methodology for expansion.
Dolgov broke down the future scaling challenges into three dimensions:
1. Core Technology: Fifth-generation hardware and continuously improved software aim to build a more general and robust 'virtual driver' capable of adapting to more cities and environments.
2. Evaluation and Deployment: Establishing an efficient process, framework, and simulation system so that engineers can focus on improving the system, while new versions of the 'driver' software can be rapidly and safely deployed to the entire fleet like an assembly line.
3. Product, Commercial, and Operational Excellence: Through the Phoenix project, accumulate hands-on experience in all business aspects, including user behavior, operational efficiency, and partnerships (including with various levels of government).
Dolgov emphasized that the value of the Phoenix project lies not only in proving the technology but also in 'forcing ourselves to learn all those hard lessons,' thereby preparing for rapid 'copy-paste' expansion to other cities in the future. He mentioned that Waymo has tested in over 25 cities, including San Francisco and Michigan (snowy environments), but the expansion of large-scale commercial services will be built on a comprehensive digestion of the Phoenix experience. Additionally, Waymo launched an autonomous trucking project called Waymo Via in 2017. Dolgov noted that trucks and passenger cars share significant commonality in underlying technologies (perception, prediction, planning, simulation, ML infrastructure), and the results of investing in core technology can naturally transfer to different product lines, though specialized adaptation is needed for trucks (e.g., sensor layout, high-speed long-distance conditions).
| Position | Analyst Stance (Bullish/Risk Warning/Neutral) | Key Data |
|---|---|---|
| Waymo | Bullish | Raised $3.2 billion in external financing; operates Waymo One paid service in ~50 sq mi area in Phoenix; 5th-gen hardware: 29 cameras, 5 lidar, 6 radar; tested in over 25 cities. |
| Jaguar I-PACE | Neutral (mentioned as the vehicle equipped with 5th-gen hardware) | 5th-gen autonomous driving hardware system deployed on this model. |
| Tesla | Risk Warning (implicit comparison) | The author (Dolgov) rebutted Musk's view that lidar is a 'crutch,' emphasizing the importance of multi-sensor fusion for safety. |
1. Waymo's Strategic Choice (Dmitri Dolgov): In 2013, the team decided to abandon Level 3 highway driver assistance and fully commit to building a fully autonomous 'virtual driver.' They believe this was an 'absolutely correct decision' because the technical paths and system designs of the two are fundamentally different.
2. Core Capability vs. Trade-offs (Dmitri Dolgov): Progress in autonomous driving does not lie in making difficult trade-offs between 'cautious' and 'aggressive,' but in shifting the entire system's precision-recall curve upward, thereby achieving safe, efficient, comfortable, and confident driving simultaneously.
3. Necessity of Multi-Sensor Fusion (Dmitri Dolgov): Lidar, cameras, and radar complement each other in different physical properties. Abandoning any one sensor is 'handicapping oneself.' The cost and aesthetic issues of lidar are not insurmountable; with technological advancement and mass production, these barriers are disappearing.
4. The Engineering Philosophy of 'Machine Learning Everywhere' (Dmitri Dolgov): Waymo's system applies machine learning deeply not only in perception but also in prediction, planning, simulation, and almost all other areas. Breakthroughs from NLP, such as Transformers, are being used to model driving behavior as a highly structured sequence problem.
5. Falsification Conditions and Scaling Methodology (Dmitri Dolgov): Waymo's operations in Phoenix are not the final goal but a 'stress test' aimed at completing a 'complete loop' from technology to product, business, and operations. Future scaling expansion ('copy-paste' to other cities) will be built on distilling all the hard lessons from this loop.