Recently, the first Ideal Home Technology Day was held in Changzhou Intelligent Manufacturing Base, LI. Under the guidance of "dual-energy strategy", LI has made important breakthroughs in the research and development of smart space, intelligent driving and high-voltage pure electric platform. The ideal SS intelligent space with Mind GPT as the core and the ideal AD intelligent driving which will open the city NOA internal test this month have entered the era of large model at the same time. The ideal 800V high-voltage pure electric platform supporting 5C charging will open the 5G era of charging speed.
The ideal AD Max 3.0, which entered the era of big model, is growing and iterating at an unimaginable speed in the past. At the product level, the urban NOA function that does not rely on high-precision maps will be delivered to the internal test users in Beijing and Shanghai, so that users can have an ever-evolving "AI driver". In the second half of the year, the commuting NOA function will be opened to users, so that users can have their own "exclusive elevator", making commuting easier and more convenient. On the technical level, AI big model technology is used to deal with all the challenges in the scene, and all the problems can be efficiently iterated and finally solved.
Perception algorithm: like people, it does not depend on maps and recognizes everything.
The core of not relying on high-precision maps is to use BEV model to perceive and understand the road structure information in the environment in real time. Through a lot of training, the current BEV model has been able to generate stable road structure information in real time at most roads and intersections.
For complex intersections, the method is to use the self-developed NeuralPriorNet network (NPN network for short) to extract NPN features of intersections in advance. When the vehicle drives to the intersection again, the previously extracted NPN features are taken out and fused with the BEV feature layer of the vehicle-side perception model, and a perfect perception result is obtained.
The NPN feature is a bunch of neural network parameters, from which humans can’t directly understand the complex intersection shape, but the large model can. Compared with high-precision maps, NPN features have more information and higher confidentiality. It uses network model instead of artificial rules to understand and use environmental information.
It is not enough to understand the intersection. "AI drivers" have to understand the traffic rules of traffic lights at intersections, which is another difficulty of urban roads. The mainstream approach is to establish a set of rules and algorithms for traffic intentions of traffic lights and roads, but the ideal is to choose a large model to solve it.
An end-to-end TrafficIntentionNet network (TIN network for short) is ideally trained. There is no need to set any rules artificially, or even to identify the specific location of traffic lights. As long as the image and video are input into the TIN network model, the network can directly give the result of how the vehicle should go now — — Turn left and right, go straight or stop waiting. By learning the reaction of a large number of human drivers to the change of traffic lights at intersections, the TIN network model is trained, and good results are obtained.
As shown in the following figure, at the intersection, the TIN network gives the probability of different traffic intentions at the intersection in real time according to the input video image, and the maximum probability value is the actual use intention, which is completely consistent with the indication of the signal light. When the red light is on, the network output waiting probability value is the maximum, and the vehicle keeps waiting. When the green light is on, the left turn probability of the network output increases to the maximum, and the vehicle starts to turn left. Although the result looks simple, the technology behind it is very advanced.
In the face of common obstacles that may appear on the road, such as construction roadblocks, scattered objects, goods protruding from the back of trucks, etc., Occupancy network is used to accurately identify their boundaries and types. In recent months, the ideal Occupy network has been iterated wildly, and a lot of training miles have been "fed", and the content and accuracy of recognition have been greatly improved.
Regulation algorithm: free control like a human.
In order to make "AI drivers" make reasonable judgments like human drivers in driving decisions and trajectories, it is ideal to apply imitation learning method to the planning and control algorithm, and train a large number of drivers’ driving behaviors, so that the decision-making and planning of urban NOA can make judgments more like human drivers on the premise of ensuring safety and conforming to traffic rules.
For example, when the vehicle needs to turn right, according to the traffic rules, you can choose either of the two lanes after turning right to merge. However, by observing a large number of human driving trajectories, it is found that more than 90% users will take the right second lane instead of the right first lane, because the safety and efficiency of the right first lane are not as good as that of the right second lane directly, and the turning radius of the right second lane is larger, the turning process is more stable, and the family will be more comfortable. Therefore, the final result of the model’s learning and training at this intersection is also inclined to take the right two lanes.
Training platform: continuous evolution like human beings.
The evolution of large-scale model needs a powerful basic training platform to complete fast and efficient training and iteration. Ideal started the construction of training platform very early. Up to today, it has grown into an autonomous driving training cluster with 1200 PFLOPS computing power, and the mileage of autonomous driving training has exceeded 600 million kilometers.
Urban NOA open logic
Whether a city can open the use of NOA mainly depends on the completion of NPN characteristics of complex intersections in the city. By counting the R&D platforms covered by NPN in cities, we can see the coverage of NPN now, and each point above represents a complex intersection that needs to make NPN features. Among them, green represents the NPN feature of the intersection, which has passed the test and verification and is available, red represents the NPN feature but needs to be verified, and gray represents no NPN feature. At present, there are many red NPN characteristic points all over the country, and there are many green NPN intersections in Beijing and Shanghai where early birds and test cars appear frequently. Next, after the early bird users join, these spots will turn green faster and faster, which means that more and more cities are open.
The commute you need more, NOA
Commuting to and from work is usually the most tiring time of the day and the time when driving assistance is most needed. If this road can be opened by NOA, it will solve the big problem. Therefore, it is ideal to introduce commuter NOA products that everyone needs more. With commuting NOA, you don’t need to wait for the NPN features of the whole city to be trained. You just need to set your own commuting route and learn NPN features by car, and you can use NOA functions on this route after learning. If you commute every day, the simple route can be activated within one week, and the more complicated route is expected to be enough to complete the training in 2-3 weeks.
It is estimated that commuting NOA can cover more than 95% of the commuting scenes of ideal car owners. During the use of NOA for commuting, each model will continue to be iteratively trained, and the better it is, the better the experience will be. With commuting NOA, it is like having your own "exclusive elevator" on the way to and from work, which is definitely the NOA function you need now.
This month, the ideal will open the NOA internal test of cities in Beijing and Shanghai, and early bird users can take the lead in using the NOA function of cities. In the second half of the year, it is ideal to open the commuting NOA function and more NOA areas in cities, so that every early bird user can use NOA navigation to assist driving during daily commuting.
All these are the advanced achievements brought by AI big model technology. Through the leading technical architecture and excellent iterative efficiency, AD Max 3.0 platform will gradually meet the needs of all cities and users to use NOA functions on urban roads. When commuter NOA and urban NOA expand on a large scale, it will be a "just-needed configuration" for mid-to high-end cars that can’t be sold well without equipment. (Photo courtesy of LI)
