Recently, with the rise of generative AI technology, many new car-making forces are exploring new methods of visual language models and world models. End-to-end intelligent driving new technologies seem to have become a common research direction. Last month, Li Auto released the third-generation autonomous driving technology architecture of end-to-end + VLM visual language model + world model. This architecture has been pushed to thousands of people for internal testing. It personifies intelligent driving behavior, improves the information processing efficiency of AI, and enhances the ability to understand and respond to complex road conditions. Li Xiang once said in a public sharing that in the face of rare driving environments that are difficult for most algorithms to identify and process, VLM (Visual Language Model) can systematically improve the capabilities of autonomous driving. This method can be achieved theoretically A breakthrough.
The new generation of autonomous driving systems has greatly increased the upper limit of capabilities - allowing AI to deal with many situations that were difficult to solve in the past, and also lowered the threshold - reducing the need for the size of technology R&D teams, and is expected to allow more people to drive in the near future Get a vastly improved experience in the future.Since the second half of last year, Ideal began to adjust its strategy and change its trajectory. In February this year, in the DriveVLM paper submitted by Tsinghua University's Cross-Information Research Institute and Li Auto, researchers applied the visual language model (VLM) that has recently emerged in the field of generative AI and demonstrated extraordinary capabilities in visual understanding and reasoning.
In the industry, this is the first work to propose an autonomous driving speed system. Its method fully combines the mainstream autonomous driving pipeline and a large model pipeline with logical thinking, and is the first to complete the large model work of end test deployment ( Based on NVIDIA Orin platform).
DriveVLM systemDriveVLM consists of a Chain-of-Though (CoT) process with three key modules:
These modules correspond to the perception, prediction and planning components in the traditional autonomous driving system process. The difference lies in their ability to handle object perception, intention-level prediction and task-level planning, which have been extremely challenging in the past.
Technical verification
Ideal verification technology is effective in long-tail scenarios:
Practical application
Li Auto’s end-to-end model and VLM model run in real time:
In complex cities In the scenario, VLM plays a role in situations where decision-making is impossible and delivers decision results and trajectories to the end-to-end model.
End-to-end approach
The end-to-end approach has become a technological watershed, marking the beginning of the real use of AI.
The new generation AI model
The new generation AI model can serve as the question maker:
Computing power challenge
車両側での VLM などのモデルの展開は、次のようなコンピューティング能力の課題に直面しています:
競争の見通し
Tesla FSD は間もなく国内スマートドライビング分野への参入 新たな競争ステージへの参入:
The above is the detailed content of L3 will be launched in the first half of next year at the latest: ideal end-to-end autonomous driving and greatly improved performance. For more information, please follow other related articles on the PHP Chinese website!