In the autonomous driving technology sector, Tesla’s FSD and Waymo’s robotaxi have long held their respective territories. However, chip giant NVIDIA is making a strong push with the “Alpamayo” system, attempting to reshape industry rules through open AI models and extreme simulation technology. Recently, NVIDIA CEO Jensen Huang personally rode in a Mercedes-Benz CLA sedan equipped with the MB.Drive Assist Pro system and completed a 22-minute全程零接管 test drive in downtown San Francisco. Huang excitedly stated that this marks the arrival of “the ChatGPT moment for physical AI.”
San Francisco Street Test: A Shocking 22-Minute Record of Zero Takeovers
The test was conducted by Xinzhou Wu, Vice President of NVIDIA’s Automotive Division, driving a test vehicle equipped with the Alpamayo system. The vehicle was equipped with the MB.Drive Assist Pro system jointly developed by NVIDIA and Mercedes-Benz, traveling from Woodside, California to downtown San Francisco. The test was conducted during peak hours with extremely heavy traffic, which was undoubtedly a huge challenge for the autonomous driving system.
From the video footage of the test drive, Jensen Huang and Wu Xinzhou appeared quite relaxed inside the vehicle. The Mercedes successfully navigated a series of complex everyday road obstacles during the 22-minute journey, including sections under construction, vehicles illegally parked side by side, and extremely narrow lanes squeezed by orange traffic cones. NVIDIA spokesperson Jessica Suarez later confirmed that there was no human intervention throughout the entire journey, achieving 100% autonomous driving.
Those who want to see the complete footage of Alpamayo’s actual performance during this road test can watch this video:
Decoding Alpamayo: The Vision-Language-Action (VLA) Model Revolution
Alpamayo is not just another autonomous driving software – it is a series of open Vision Language Action (VLA) models launched by NVIDIA. The core concept of these models is to give vehicles “reasoning capability,” not just reactive capability. According to NVIDIA’s technical blog,ExplanationAlpamayo combines open AI models, the AlpaSim simulator, and extensive physical AI datasets to build “thinking” autonomous vehicles.
NVIDIA developed this open-source toolchain hoping to enable all car manufacturers to easily give vehicles human-like decision-making capabilities, just like using ChatGPT. This contrasts sharply with Tesla’s current closed, self-use-only technology approach (though it does require using NVIDIA’s system).
Technical Path Rivalry: End-to-End AI vs. Classical Engineering’s Vertical Integration
Jensen Huang emphasized in the video that NVIDIA’s solution is “one-of-a-kind” because it perfectly combines end-to-end AI models with the traditionally engineered “classic” tech stack. The end-to-end model enables a more natural driving style (handling speed bumps, lane changes), while the classic tech stack ensures absolute enforcement of safety rules. Pure end-to-end models make safety verification difficult, while NVIDIA’s hybrid system can maintain flexibility while meeting the most stringent industrial safety standards.
Facing the hundreds of millions and billions of real-world miles accumulated by Waymo and Tesla, NVIDIA sees things differently. WU Xin-zhou put it bluntly: “The real infrastructure is simulation.” NVIDIA leverages Neural Reconstruction (NuRec) technology to recreate scenarios 1:1 from real-world data, and uses data augmentation to generate millions of extreme scenarios in the virtual world. It’s like having AI complete tens of thousands of hours of intense training in a simulator before ever setting foot in a driving school.
The future of driving: a traffic rulebook and 20 hours of training?
NVIDIA’s perspective on autonomous driving: when humans teach children to drive, there’s no need for them to log thousands of kilometers first. Instead, we teach them to understand traffic rules, then let them practice on the road for about 20 hours. NVIDIA’s VLA model is working toward exactly this goal: by leveraging visual and language understanding, the model can fundamentally “comprehend” driving logic, freeing it from reliance on massive amounts of real-world driving data. As for which of the three approaches—Tesla, Waymo, or NVIDIA—holds the correct answer to autonomous driving, we may find out on the roads in the not-too-distant future.
Source: KOCPC Chinese