Xiaomi's autonomous driving technology capabilities have received the latest international certification.
At the autonomous driving challenge held at the top AI conference CVPR, Xiaomi won the second place in the highly anticipated map-free track.
Advertisement
This is also the first global research recognition Xiaomi has received since it entered the car manufacturing industry and started independent research on autonomous driving.
Compared with other players, Xiaomi can be said to have gotten on the bus first and bought the ticket later, mass production first and then demonstrated its research level.
The International Autonomous Driving Challenge is a competition under the AI vision top conference CVPR 2024. It has a total of 7 sub-tracks and attracted hundreds of teams from more than 10 countries to participate. It can be said that there are many masters and a fierce competition.
Advertisement
Xiaomi made its debut and achieved good results. However, the honor also belongs to Beijing University of Posts and Telecommunications. This time, the Xiaomi team, together with Beijing University of Posts and Telecommunications, proposed a new method for understanding road topology and jointly won the challenge.
It is worth mentioning that the well-known Tier 1 Bosch Global only ranked fourth.The top three are all Chinese players without exception, and they may all be the eye-catching “new players” in autonomous driving.
First place: Langge Technology.
Second place: Xiaomi & Beijing University of Posts and Telecommunications.
Third place: Megvii & Beijing Institute of Technology.
Xiaomi ranked second in the world in the top autonomous driving challenge
A series of challenges are held before and after the CVPR conference, which has become a tradition. The most popular one at present is the Autonomous Grand Challenge.
This year's challenge has a total of 7 tracks:
End-to-end autonomous driving
World Model
Motion Prediction with Occupancy Networks and Optical Flow
Embodied Intelligent Multimodal 3D Visual Positioning
CARLA Autonomous Driving Challenge
Application of LLM in autonomous driving
Autonomous driving without a map
The organizers of CVPR said that the emergence of LLM and world models brings huge potential to autonomous driving, and explainable end-to-end models will be the future.
Such models can understand the world and generate scenarios that have never been actually visited.
In view of this, the meeting decided to divide it into 7 tracks to promote breaking the boundaries of today's perception, prediction and regulation.
As the key to the current autonomous driving race, the mapless driving track has also attracted much attention – it is considered to be the key technology for the larger-scale implementation of mass-produced autonomous driving.
The challenge brought together 120 teams from 10 countries, including FAW, Bosch (Germany), Xiaomi, Megvii Technology and Lange Technology, and received a total of 729 jobs.
The “no map” mentioned here also refers to abandoning the high-precision map and using the standard definition map (Standard Definition Map) instead, which can be understood as a navigation map.
△The left side is a sample of a standard map, and the right side is a sample of a high-precision map of the corresponding road
The accuracy of the precision map is usually at the meter level. It provides basic road and lane information and can assist the autonomous driving system in navigation and scene understanding.
The work of Xiaomi and Beijing University of Posts and Telecommunications has proposed a new method of using precise maps to assist in understanding road topology.
Specifically, the research team designed a compact Transformer-based architecture for encoding and integrating precision maps, utilizing intermediate lane points, improving each lane attention layer in the detection head, and dynamic decoding.
This can improve decoding performance, obtain more accurate location information, and predict scene changes during driving.
After fierce competition, this work finally won the second place in the map-free autonomous driving track.
Who is the top player on the track?
A little-known company: Langge Technology.
Founded in 2021, Longo Technology is no ordinary startup.
Public industrial and commercial information shows that the actual controller behind the company is Li Shufu, who holds a controlling stake of over 96%. Lango’s chairman Liu Jinliang is a director of Geely Group.
It mainly engages in intelligent driving maps, multimodal navigation and intelligent driving data services. It is currently the only automotive intelligent service provider in the industry with both Class A surveying and mapping qualifications for OEMs and navigation electronic map production (this qualification comes from its subsidiary Hubei Ecarx Technology).
Previously, we developed a large Zhoutian multimodal map model, which was used as a basis for this competition. We optimized it for the competition tasks, utilized the road topology prior provided by the standardized map, introduced deep auxiliary supervision, and proposed a hybrid time series fusion module to achieve the best road topology structure recognition performance.
With their success at this year's CVPR, both Longor and Geely have more reasons to be confident in intelligent driving.
It's time to say goodbye to Mobileye dependence.
Other tracks of the CVPR Autonomous Driving Challenge
In addition to the Wutu autonomous driving track, there are many familiar names on the performance lists of other tracks.
In the remaining six tracks, Huawei ranked on the list in three tracks, winning the second place in the World Model Track, the third place in the CARLA Autonomous Driving Challenge MAP Group, and the fourth place in the Occupancy Network and Optical Flow Track.
The last thing I want to introduce is the hottest and most noteworthy end-to-end track at the moment. The first place is the giant NVIDIA, which has joined forces with Fudan University and many other institutions.
A novel architecture called Hydra-MDP is proposed, which uses a multi-teacher teacher-student paradigm for knowledge distillation, integrating knowledge from rules and people. This enables the model to learn different trajectories and improves generalization capabilities in different driving scenarios.
It is worth mentioning that the perception network in this work uses data from cameras and lidars, which is a fusion perception solution.
Following closely behind NVIDIA, the team that won second place in the end-to-end autonomous driving challenge was Suzhou-based startup Zero-One Auto, which was founded by TuSimple co-founder Huang Zehua. The winning team members included TuSimple veteran Wang Panqu and others.
The work submitted by his team combined the multimodal visual language model (VLM) with an end-to-end architecture and applied it to autonomous driving, achieving good performance.
Therefore, the CVPR International Autonomous Driving Challenge can be said to have brought together heroes from all walks of life, including international giants and promising start-ups.
AI conferences have always been the first place for the latest and most powerful technologies to emerge, and challenge competitions are the best battlefield for many start-ups to become famous.
This year, Xiaomi Auto made its debut and made a splash.
One More Thing
However, the AI top conference award did not become a new hot search for Xiaomi Automobile. Instead, the second car that Xiaomi revealed, the Xiaomi SUV, dominated the attention.
Its road test spy photos have been exposed, and it still has a long front and a relatively low-lying appearance.
Even with heavy camouflage, netizens still gave this car high praise, believing that the Xiaomi car design team lived up to its reputation.
Continuing the successful experience of Xiaomi SU7…
But soon, a large number of netizens and bloggers pointed out that the appearance of this car seemed to have been seen somewhere before… Friends who know more about cars further posted this picture:
△Image source: Weibo@科技头条
That’s right, it’s Ferrari’s first SUV, the Purosangue… But beauty is in the eye of the beholder. Some netizens think there are still some differences, and it’s more likely to be similar to the relationship between Xiaomi SU7 and Porsche Taycan – a tacit understanding.
However, even if there are different opinions, the final consensus is the price. Xiaomi's SUV will definitely not be priced at 5 million like the Ferrari Purosangue.
Xiaomi moment is about to come to SUV models: 5 million appearance design, 5 million hardware configuration, 5 million premium experience…
The selling price does not exceed 500,000.
Friends who want to buy a Ferrari Purosangue, wait a little longer?