The global competition in autonomous driving is intensifying. In February, Waymo operated approximately 3,000 vehicles, and by the end of June, its cumulative distance driven without a driver surpassed 270 million miles. China's expansion is also rapid, with Baidu's Apollo Go exceeding 240 million kilometers in fully autonomous driving and Pony.ai operating 1,975 robo-taxis globally by the end of June.
So, where do we stand? In Seoul, the number of autonomous taxis increased to 19 as of August, but they are still operating with test drivers in limited areas and time slots. While various autonomous driving demonstrations are taking place across the country, large-scale commercial services without a driver are still in the early stages. The gap between us and the U.S. and China, where thousands of vehicles are operating daily in commercial services, must be acknowledged.
More importantly, the difference lies not just in the number of vehicles but in the data they generate. In AI autonomous driving, vehicles collect data, which is used to train AI, and the improved AI is redeployed back into the vehicles. As performance improves, the operational range and number of vehicles increase, leading to even more data generation. This cycle of data collection, learning, validation, and deployment is known as the 'data flywheel.' Once it starts turning effectively, the speed of learning and improvement accumulates, widening the gap between early and late entrants over time.
Thus, our most urgent task is not to wait for perfect autonomous driving technology. Instead, we must deploy a significant number of vehicles on real roads and quickly initiate the flywheel of data collection, learning, validation, and redeployment. Based on my experience, this flywheel does not start turning after the technology is perfected; rather, the technology is refined during the process. Errors and interventions discovered in real environments, along with numerous edge cases, feed back into learning, allowing AI to evolve. Starting late does not mean we have no chance; the greater risk lies in continually delaying our start.
However, the real-world data flywheel is a necessary but not sufficient condition. Simply chasing the billions of kilometers of real-world driving accumulated by early entrants will not close the gap. Latecomers need a strategy that not only initiates the flywheel but also significantly accelerates its rotation.
One method is to combine the real and virtual worlds. By utilizing simulations and generative world models based on real data, we can enhance the quantity and diversity of situations that AI can experience, thereby increasing the 'density of experience.'
Waymo has driven hundreds of millions of miles on real roads but has already surpassed 20 billion miles in the virtual world. The world model released by Waymo this year generates rare scenarios that are hard to encounter on real roads and simulates what might happen if vehicles made different choices in the same situation. Simulations have evolved into closed-loop learning environments where AI can repeatedly learn from actions and outcomes to improve policies.
In the robotics field, these changes are happening even faster. The Sim-to-Real transfer method is gaining traction, where numerous virtual robots are trained simultaneously, experiencing failures such as falling, slipping, and colliding, and then applying improved policies through reinforcement learning to real robots. Failures that are difficult to replicate in reality due to cost and risk can be transformed into valuable learning assets in the virtual world.
Recently, the world has been captivated by physical AI and robotics. However, autonomous vehicles are essentially 'robots on wheels.' They share a technological trajectory by recognizing their surroundings with sensors, predicting changes in the environment, making decisions, and learning from the outcomes. They share a significant portion of the core technology stack, including data, AI models, world models, simulations, reinforcement learning, AI semiconductors, safety verification, and operating systems.
In this context, dreaming of becoming a robotics powerhouse without having the foundational technology to train and operate wheeled robots on a large scale may be unrealistic. Autonomous driving is not a separate industry but the most practical testing ground for accumulating and validating physical AI development capabilities in real environments.
Our time is limited. The data flywheels of global leaders are spinning rapidly even now. What we need is not to wait for perfect preparation but to ignite a large-scale flywheel immediately, connect the real and virtual worlds to increase its rotation speed, and extend the technologies gained here to robotics and manufacturing.
Autonomous driving is a competitive edge for the automotive industry and a strategic foothold that South Korea must secure to become a powerhouse in physical AI and robotics. What we need to build now is the foundation for physical AI that will evolve the next-generation industry, with AI autonomous driving serving as the first platform built on that foundation.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.