Part 2/9:
The research demonstrates that self-play can effectively develop robust and naturalistic driving behaviors, showcasing the potential for the technology in autonomous vehicles. By simulating driving scenarios at an unprecedented scale, Apple claims to have achieved remarkable results, using their software package called Giga Flow to synthesize and train self-driving agents.
Unprecedented Scale of Simulation
The paper reveals that Apple’s model processed an astonishing 1.6 billion kilometers (about 1 billion miles) of simulated driving in just ten days on a single GPU node. This capability allows for the synthesis of 42 years of subjective driving experience every hour, enabling the agents to learn how to navigate complex traffic scenarios without relying on real-world human data.