Greater Bay Area Science Forum｜Professor Huang Hui of Shenzhen University: Intelligent graphics perception solves the problem of 'stuck neck' of robots

Southern Finance All-Media Trainee Reporter Ma Jialu Nansha Report

The further development of robot intelligence is restricted by the lack of real three-dimensional dynamic environment capabilities. "On May 20, Huang Hui, a national leader in scientific and technological innovation and dean of the School of Computer and Software at Shenzhen University, introduced her research on intelligent graphics perception at the female scientists sub-forum "Looking for Her in the Crowd" of the Greater Bay Area Science Forum As a result, UAVs are used to conduct fully automatic and autonomous exploration of three-dimensional space, and higher-precision modeling can be achieved with a smaller data volume. Currently, this achievement is in an international leading position and can be used in fields such as digital twins and intelligent robot manufacturing. widely used.

Smaller data size, higher accuracy

The Institute of Physics of the Chinese Academy of Sciences once popularized the definition of digital twin: digital twin, the English name is Digital Twin (digital twin), also known as digital mapping and digital mirroring. Digital twin is to create a digital "copy" based on a certain device or system. Simply put, it is to transform physical objects in the real world into digital virtual experiences.

In early 2023, the "Digital Twin City Industry Map Research Report (2022)" released by the China Academy of Information and Communications Technology observed that the development stage of my country's digital twin industry is in a growth period, the digital twin city construction market is active, and market entities are constantly expanding. Digital twins can greatly improve the construction efficiency and real sense of the "Metaverse" and are an important part of the "Metaverse". With the popularity of the concept of "Metaverse", the cutting-edge technology of digital twins is becoming more and more familiar to people. .

In the past, data twins required a lot of manpower and material resources to collect data, and then used a lot of calculations to build models. Huang Hui pointed out that the cost of data collection is high, the timeliness is low, and the objects of 3D modeling are unstructured and uncorrelated, which are the pain points that make it difficult for current digital twins to achieve intelligence, modularity, lightweight, structure, and generalization.

In response to the above problems, in the past ten years, Huang Hui has focused on "intelligent graphics and perceptual computing" to overcome difficulties, and has led the proposal of Youshi precise photogrammetry technology to fully automatically collect urban three-dimensional images with millimeter-level high-definition resolution through portable drones. Snapshot, the entire process is continuously planned and crowdsourced from end to end, using the least viewing angles to achieve maximum coverage, reducing the amount of large-scale urban scene data by 200 times, while ensuring a geometric detail loss of less than 0.6 meters, making data collection equipment Investment can be reduced by 70%. Achieve accurate mapping of multiple information and benign metabolism. This technology has been widely used in a single project of close-in high-precision urban three-dimensional aerial survey, and the coverage area maintains an international leading position in the world.

Take the modeling of approximately 2,200 square kilometers of Shenzhen City as an example. In practice, modeling faces the challenges of difficult airspace coordination, numerous altitude restrictions, frequent weather effects, poor data quality, and long collection time. If the traditional modeling method is used, it will take 2 years to collect data of 60 million images and build a 60TB model, with a total cost of 150 million. Real-time updating of the three-dimensional basis of the city twin cannot be achieved because the acquisition method has a long cycle and low frequency. After optimization of Ushi's precision photogrammetry technology, it is possible to build a 6TB small model using 20 million photo data in 8 months without reducing accuracy, and the total cost is reduced to 60 million, a reduction of 60%.

Digital twins have important value and significance in fields such as swarm intelligence, driverless driving, smart cities, homeland security, and industrial manufacturing. "Huang Hui said that Youshi's precision photogrammetry technology uses drones to conduct fully automatic and independent exploration of three-dimensional space and realizes cloud computing, which greatly reduces the cost and threshold of three-dimensional urban reconstruction. It can be widely used in high-precision intelligent driving maps in the future. Construction, real-life three-dimensional navigation, high-precision urban management information platform and other fields.

The lack of three-dimensional perception ability "stuck the robot's neck"

The "Digital Twin City Industry Map Research Report (2022)" proposes that in recent years, artificial intelligence technology has gradually been applied to the field of modeling, and picture modeling and video modeling have become future development trends. According to Huang Hui, the application of artificial intelligence to assist robot research may improve the efficiency of data analysis and help find more efficient three-dimensional modeling solutions. She revealed that she and her team are studying methods to combine intelligent graphics and artificial intelligence to further advance research on intelligent robots.

A challenge for intelligent graph perception is the lack of data available for large-scale training, compared to ChatGPT. Huang Hui said that the original data used by ChatGPT has basically been manually screened and annotated, and is "clean". However, when this extends to three-dimensional space, the difficulty will increase significantly. An item can be photographed 1,000 times, but all 1,000 photos may not fully and accurately describe the item. Re-stated as follows: This means that when training an artificial intelligence model to process small-scale three-dimensional spatial data, it will need to process a large amount of data.

The robot’s intelligence is stuck because it lacks real three-dimensional dynamic environment capabilities. "Why is it so difficult? Huang Hui used the analogy of a common sweeping robot to explain: It is not easy now to make a sweeping robot move on a flat surface, sense and avoid obstacles, and prevent itself from getting stuck. Compared with it, intelligent graphics perception It is necessary to be able to judge the shape of objects in three-dimensional space, and then judge the functions and dynamic and static relationships of objects. It is necessary to give robots embodied intelligence that can have more advanced perception, and be able to truly face and change the real three-dimensional complex dynamic environment, and "see" like humans. That is to understand, what you see is what you get”, make decisions and actions.

After more than 20 years of basic research, Huang Hui led the team to try to unify the entire chain of environmental perception, geometric modeling, semantic understanding, and autonomous decision-making. However, compared with artificial intelligence technologies such as ChatGPT, which have made significant progress in natural language understanding, there are still many key scientific issues that have not been solved, and "there is still a long way to go."

For more content, please download 21 Finance APP

The above is the detailed content of Greater Bay Area Science Forum｜Professor Huang Hui of Shenzhen University: Intelligent graphics perception solves the problem of 'stuck neck' of robots. For more information, please follow other related articles on the PHP Chinese website!