Turing Achieves Japan’s First Public-Road Autonomous Driving With a VLA Model

Turing announced that it had achieved real-time autonomous driving on public roads using a VLA model for the first time in Japan. The company independently developed a VLA model with about 2 billion parameters, enabling inference and control at 10 Hz. It also released the causal reasoning dataset “RACER” and the image tokenizer “DriveTiTok.”

For companies and researchers developing autonomous driving, the key point is that Turing demonstrated that an approximately 2 billion-parameter VLA model could run inference at 10 Hz on an in-vehicle computer and control steering, acceleration, deceleration and braking in real time on public roads. The company also released a causal reasoning dataset with explanations for decisions and an image tokenizer that compresses driving footage to about one-hundredth its size while retaining information needed for driving decisions. These provide concrete resources for preparing training data and validating systems for in-vehicle use.