How to control costs with Mapfree smart driving solution?

Since this year, light maps have gradually become a popular keyword. High-precision maps are gradually being replaced by light maps due to limitations such as map freshness, maintenance costs, and coverage.

Yuanrong Qixing is one of the earliest companies in the industry to lay out light map solutions. By reconstructing the technical framework and conducting a large number of general tests, it found a balance between safety, efficiency and somatosensory, and finally made a The cost of a set of solutions is only one-tenth of the cost of traditional high-precision maps, making mapless solutions possible.

According to Liu Xuan, the technical route of Yuanrong Qixing is not exactly the same as that of Tesla . It does not require the vehicle to remember the specific data of a certain intersection. Instead, it is hoped that the model itself can learn to think like a human and drive while driving. Each data in the operation may be within a certain range, and then the system completes the operation autonomously.

According to the plan, Yuanrong Qixing will provide traditional L2+ assisted driving and high-speed NOA solutions this year, and will provide pure visual products for urban NOA next year.

Mainly based on real-time perception, with SD Map as reference input

Q: Although everyone in the industry is talking about maps, most car manufacturers and suppliers still use the so-called SD Pro or Lite maps as their core. Yuan Rong can achieve high-level assistance using only SD Map. What is the technical support behind this?

A: Whether it is SD Pro or HD Lite, it still contains high-precision map information. It basically uses high-precision maps to cut down and provide some richer information, which can help the algorithm do better. However, issues related to the update of high-precision maps and the cost of updates are still unavoidable. If this problem can be solved at an algorithmic level, then it may no longer be a particularly large strong dependency.

Demapping is definitely a future development direction because it can reduce maintenance costs. The maintenance and update part is essentially similar to the manual part of artificial intelligence . If this problem can be solved from an intelligent perspective, the cost can be reduced a lot.

We first started research and development in 2020, and the technology is relatively mature. Therefore, based on our experience, we believe that the dependence on high-precision maps can be eliminated.

Q: The research and development focus in the industry has shifted from relying on high-precision maps to eliminating high-precision maps. Is it a simple cost issue or is the development of high-precision maps unable to keep up with the speed of implementation?

A: In fact, in the industry, including OEMs, some start-up companies also considered high-precision maps earlier, but the main reason may be that the technology was not mature enough in the past few years, so they did not invest in this direction. Later, after Tesla and Yuan Rong made it, everyone found that this technology was feasible. Therefore, this year everyone believes that this technology is close to a breakthrough point, and more and more people recognize this technology.

Q: If the picture-free solution is mass-produced, does it need to be generalized for every city? Due to the re-awareness of graphless solutions, does this mean that the system requires strong data closure?

A: First of all, because most roads in cities are relatively regular, the significance of generalization is not particularly great. Basic vehicles can be learned after driving once on a relatively regular and simple road section, so they can also be used in other cities.

The difficulty is that for those irregular corner cases, such as road construction or elevated scenes, where people may drive the wrong intersection just by looking at the map, we need more data. Therefore, complex roads in cities, including some rural roads, require a large amount of data for generalization, so of course it must rely on a strong data closed-loop system.

This system is very comprehensive. It includes perception, positioning, planning, etc., and has a complete set of interactive processes. We need a lot of data, especially if we want to transition to pure vision in the future, we will need more data. So how to find useful information from massive data and make this algorithm iterative better? It requires very high capabilities and is more difficult than the previous traditional solution.

Q: Under this model, how can OEMs and integrators cooperate efficiently to open up the data closed loop?

A: The ownership of the data generally belongs to the OEM, but the OEM will share the data with us for joint construction, allowing us to update the algorithm model to adapt to more road conditions and achieve better performance. Ultimately, we hope that this product can be continuously iterated and upgraded services provided to users. In addition, these data will be on a controllable and trustworthy cloud platform. We only need to enable it to train the model.

Q: What are the requirements or changes in the Mapfree solution for positioning devices such as combined inertial navigation?

A: Our current combined inertial navigation products include IMU and GNSS, which are about a hundred yuan level. There is no need for differential positioning, it only needs to be able to provide positioning accuracy of about ten meters. Then it can meet safety needs through algorithms, which can also reduce certain costs for car companies.

Q: How can the mapless system make good use of SD maps during smart driving?

A: For the application of SD map, we mainly use it as a reference information, more based on actual perceived information, and then use the navigation map as a reference, so that it can learn a more human-like value behavior .

Q: How much cost pressure will the car manufacturer reduce with pictures, light pictures and no pictures at all?

A: A pure SD car costs about dozens of yuan a month, but an HD car may cost a few hundred yuan a year, and SD Pro or HD Lite may cost between 100 and 200 yuan a year. piece. But the current problem is that the city data covered by HD and SD Pro is not enough. There are currently only six cities. If more cities are opened in the future, the cost will inevitably increase, so this cost is not necessarily a fixed value.

Q: Compared with the cost of light maps and high-precision maps, is there not much difference in the costs incurred during the collection process?

A: Yes, and it collects data. Subsequent updates also need to go through the drawing review period, so the frequency of updates may not be that fast. If it relies entirely on high-precision information and the update frequency is not high enough, it may also have a certain impact on the performance of the algorithm.

Q: What is the rhythm of urban generalization? How long does it take to generalize from one city to another in the early stage?

A: Currently, test vehicles are collecting and testing generalized data in 16 cities. Because it does not rely on high-precision maps at all, you only need to put the car into the city for testing. Because of the recognition of information such as traffic lights on the road, the model itself has undergone certain robustness and generalization training. Secondly, the most challenging thing in different cities is not the perception results, but the complexity of the road structure.

In fact, the so-called generalization test is more about collecting static information of the road network so that it can learn difficult and challenging scenarios on different urban roads. Because it is Mapfree, it does not rely on navigation maps, so there is little need to prepare in advance, so generalization requires less time.

Not exactly the same as Tesla, more focused on model learning

Q: How does Yuan Rong’s unpictured solution deal with the access requirements for ICV/L3 functional safety?

A: First of all, we have specific indicators for the algorithm itself, and we regularly compare it with our own algorithm using high-precision maps or algorithms purchased on the market. Compare the product effects, including system safety, comfort, driving efficiency, etc.

Regarding functional safety, the car companies we are currently cooperating with also have very high requirements for this. During the cooperation process, we will break down the safety needs of car companies, and then conduct system-level software and even hardware monitoring according to each module and according to the requirements. For example , we have corresponding monitoring for automotive MCU chips , and for some fault events on Orin chips, we have corresponding modules to monitor to ensure that when the system has some faults or failures, the corresponding information can meet the needs of car companies. functional safety status.

Q: When the road environment is complex and blind spots occur, can the current road environment be predicted in time? When traffic congestion creates sensory blind spots, will the effectiveness and safety of functional implementation be reduced?

A: For information about obstructions at intersections, if an object has been identified before, then within a short period of time, the perception system can track the object through multi-frame fusion perception, and we have some internal benchmarks with very high accuracy. . Unless it is a ghost probe that has no previous information, it may be more difficult, but this situation is a very difficult challenge for any algorithm.

The more difficult point is that when Mapfree recognizes the road, if it cannot identify the stop line due to traffic jam, the algorithm will use machine learning to guess the location of the stop line, based on the experience of a large number of human drivers or based on reinforcement learning. judge.

In fact, I think Mapfree brings more convenience to traffic jams or road construction situations than traditional high-precision maps.

Because if the real road and the map do not match, the navigation map will have a lot of wrong information due to not being updated in time, and the traditional vehicle may lose its direction. However, everything in Mapfree is based on what is happening on the actual road. This is the most accurate information. . So I think it is more reliable from a security perspective.