英伟达要做Robotaxi，采用端到端+强化学习｜36氪独家

Core Insights - Nvidia is expanding its business by incubating a Robotaxi project, which will be led by senior director Ruchi Bhargava, and aims to create a "technical sample" for Robotaxi technology [1][2] Group 1: Project Overview - The Robotaxi project will utilize a new one-stage technology route, employing an "end-to-end" neural network reinforced by a world model created through simulation technology, similar to Tesla's FSD approach [1] - Nvidia plans to invest $3 billion in the Robotaxi project, with expectations for it to launch in the U.S. [2] - The project is seen as a way for Nvidia to validate its full-chain engineering capabilities from GPU chips to physical AI models [2] Group 2: Market Context - The U.S. market for Robotaxi is expected to accelerate, with companies like Waymo and Tesla already making significant strides in the sector [3] - Waymo operates approximately 700 vehicles and has plans to expand its commercial operations in 2025, while Tesla's Robotaxi service has seen high initial download rates [4][3] - Regulatory changes are anticipated, with the NHTSA planning to revise existing regulations to facilitate the deployment of Robotaxi vehicles [3] Group 3: Competitive Landscape - Nvidia has faced challenges in developing advanced autonomous driving software, with past collaborations yielding mixed results compared to competitors like Momenta [6] - The company is still catching up in the L4 autonomous driving space, with significant gaps in talent, production-level algorithms, and real-world testing experience compared to leaders like Waymo and Tesla [7] Group 4: Technological Advantages - Nvidia's core strengths lie in its chip and computing ecosystem rather than direct fleet operations, with its DRIVE Thor chip offering 2000 TOPS of computing power to enhance model inference efficiency [7] - The company has substantial financial resources, with a net profit of $26.4 billion in Q2 2025, providing ample support for the long-term development of the Robotaxi project [7]