Baidu Li Zhenyu: Big models will move towards multimodality, shaping fully autonomous automotive robots

Lei Diwang Lotte October 17th

Founder of Baidu Robin Lee, chairman and chief executive officer, said at Baidu World Conference yesterday: "The future AI native applications must be multi-modal. Beyond the information world, the physical world will be reconstructed. Automatic driving is a typical application of the visual big model to reconstruct the physical world. The big model will allow Baidu's automatic driving ability to surpass the experience system, deal with complex scenes more intelligently, and achieve a broader space-time coverage."

At the 2023 "Big Model 'Reconstruction' of Intelligent Cars" forum at the Baidu World Congress that afternoon, Li Zhenyu, Senior Vice President of Baidu Group and President of the Intelligent Driving Business Group, stated that the "intelligent emergence" of big models brings breakthroughs in core abilities such as understanding, generation, reasoning, and memory, giving cars EQ and IQ, and will restructure the intelligent car industry. The future big models will also move towards multimodality, shaping fully autonomous automotive robots.

Radish Run will be getting closer to commercial profitability

In terms of smart cabin, the use of the language model will upgrade the interaction mode between people and vehicles from "command based" to "dialogue based", promoting the upgrading of the relationship between people and vehicles to the relationship between people and virtual humans. The large model will reconstruct the way people and vehicles interact, making the interaction more natural. Based on the Wenxin large models, Baidu Apollo has created a special big model technology base for the car cockpit.

The interaction between people and cars no longer requires complex button operations, and can be controlled using voice. Even in situations of various tongues, multi person commands, interweaving voices, and continuous dialogue, the intelligent cockpit can understand each person's different needs and meet them simultaneously. At present, the products supported by the Baidu Apollo intelligent cabin model will be mass-produced and equipped in brand models such as the Geely 01, Cadillac, Buick, and Geely Galaxy.

In terms of intelligent driving, the autonomous driving technology stack has been thoroughly restructured through new technologies such as Transformer and BEV, resulting in an improved sense of intergenerational perception and accelerating the maturity and popularization of pure visual solutions. Baidu stated that its Apollo pure visual high-end intelligent driving solution can be applied to high-speed, urban, parking and other global scenarios, and will achieve mass production in the fourth quarter of this year. This is the first pure visual solution to be implemented in urban scenes in China. Removing LiDAR reduces the overall cost of the vehicle and enhances market competitiveness.

Large models will also move towards multimodality and reconstruct the physical world, with autonomous driving being a typical representative of the reconstruction of the physical world by large models. The large model allows autonomous driving to surpass experiential systems, handle complex scenarios more intelligently, achieve wider spatiotemporal coverage, and shape fully autonomous automotive robots.

400Radish Run will be getting closer to commercial profitability

Li Zhenyu stated that Baidu has invested over a decade in fields such as artificial intelligence and deep learning, and has also explored the field of intelligent vehicles. The accumulation and practice of technology over the past decade are the source of confidence and confidence for Baidu Apollo. The big model will truly land fully autonomous driving, and the wave of intelligent cars will also quickly arrive.

Three Key Paths for "Reconstructing" Intelligent Cars in Large Models

In recent years, the proportion of intelligent driving in the overall purchasing factors of users has rapidly increased, with the proportion of "most important factors before purchasing" breaking through from 1.2% to 30%, becoming the core decision-making factor for users to make car purchases. The smart car market is on the eve of large-scale production, and the underlying intelligent technology of smart cars is also undergoing restructuring, allowing the era of AI native travel to arrive faster.

Li Zhenyu believes that the reconstruction of the smart car industry by large models is mainly reflected in three aspects. The language model will be upgraded from "command based" to "dialogue based" when getting on the car; By thoroughly reconstructing the autonomous driving technology stack through new technologies such as Transformer and BEV, the perception ability is enhanced to enhance intergenerational sense, accelerating the maturity and popularization of pure visual solutions; The future big models will also move towards multimodality, shaping fully autonomous automotive robots.

At the meeting, multiple integrated intelligent driving and cabin driving products were also released. The Apollo HighwayDriving Pro produced by Baidu Apollo has further evolved and released a new generation of Apollo CityDriving, upgrading its usage scenarios from closed roads to urban open roads, with functional scenarios infinitely close to the entire domain.

Baidu stated that the pure visual city navigation high-end intelligent driving product ApolloCityDriveingMax will be mass-produced and launched in the fourth quarter of 2023. At the same time, Baidu Apollo also launched the industry's first Apollo Robo Cabin integrated soft core intelligent computing platform, which is the first platform in China to achieve or even the world's first truly integrated operation of cabin and cabin on a single SOC.

At the event, Baidu Apollo and Hangsheng signed a strategic cooperation agreement, announcing that they will jointly create a new generation of cabin driving integration products based on the Qualcomm platform.

Lei Di was founded by media person Lei Jianping. If reprinted, please specify the source.