Apple releases new self-driving training model that doesn't rely on real data

Apple published a new autonomous driving research paper on its machine learning research page, which proposed a new large-scale self-game reinforcement learning framework called GIGAFLOW. This framework can effectively train general driving strategies and is highly robust. According to data, through this self-supervised training system, 1.6 billion kilometers of training can be completed every 10 days, equivalent to 9,500 years of driving experience. This training method is also very cost-effective, with a simulation cost of less than $5 per million kilometers, and no real driving data is required.