Why doesn't Tesla use a high-definition map

Editor's introduction: Almost all domestic and foreign self-driving manufacturers have invested in high-precision maps, but Tesla is an exception, what is the reason behind this? Perhaps, we can start from the characteristics of the high-precision map itself, as well as the road environment at home and abroad. In this article, the author has made an interpretation of why Tesla does not use high-precision maps, let's take a look.

We can see that domestic and foreign autonomous driving manufacturers (especially domestic map vendors) are laying out high-precision maps, and Tesla said that it is not cold, completely abandoning non-camera sensors such as lidar and millimeter-wave radar, and only using cameras for perception, which is unique in the field of automatic driving.

This caused a discussion between me and the leader, and I went back to comb through it and record my thoughts here.

First, what can high-precision maps do

The HIGH-precision map can be seen as a model of the road environment, recording the three-dimensional characteristics of the road, driving assistance information (such as lane lines, etc.) and rich semantic information (such as the type of traffic lights, etc.), and improves the "upper limit" of the automatic driving effect by becoming a high-precision positioning basemap, providing planning materials, and strengthening the ability to perceive.

1) Improve positioning accuracy (reference object of perception system)

Confirm your current position by comparing the identification results of the on-board positioning module and the perception module.

2) Provide road information beyond the visual distance, plan materials (increase the upper limit of decision-making distance, liberate computing power, and smoother driving experience)

The detection distance of the vision/perception system is limited, especially when the vehicle speed is fast, leaving the reaction time for the on-board computer to be short.

The cloud can preprocess the moving line planning based on the high-precision map, saving the computing power of the on-board computer.

Even if the local computing power is enough to correct the deviation in a short period of time, it will sacrifice the riding experience (seeing to stop and then stop, and knowing in advance to stop and slow down in advance, the difference in experience).

Limited visual range, combined with over-the-horizon information, reduces local optimal decision-making.

3) Help unmanned vehicles identify vehicles, pedestrian positions, obstacles, and road signs (give the perception quality a bottom line, so that decision-making is more reasonable)

The visual scheme will be due to the light and darkness, the color of the object, etc. will produce misidentification, the radar will be affected by fog and rain (particles in the air) to produce noise, high-precision maps do not, but the sensor found that the probability is an active object.

Compare and calibrate with the perceptual "seen" street sign markings to reduce misidentification.

It can be seen that the high-precision map mainly provides "gain" for the perception and decision-making modules to improve the safety and comfort of the system.

Just like an exam, the side that does not bet on the question, the unmanned car that does not rely on high-precision maps, or even only uses pure visual schemes, may only be able to answer 60 points, or desperately learn to get 80 points. Asked the tutor to get this side of the syllabus, with the help of high-precision maps, through other ways to bypass the visual dead spots, difficulties, get 90 points.

Second, do not want to use OR can not be used?

Tesla chose this more radical route, completely abandoning non-camera sensors such as lidar and millimeter-wave radar, and only using cameras for perception, which is unique in the field of automatic driving.

From the perspective of "first principles", even without high-precision maps, highly mature perception of "vision", unmanned vehicles can use only "eyes" to make driving decisions like human driving.

I tend to think that Tesla is not "not wanting" to use HD maps, but "can't" use HD maps (which are not cost-effective).

1) Tesla for the global market, the production and maintenance cost of high-precision maps is high, and the effect is not good

The production of high-precision maps generally includes a series of data production processes such as collection, data processing, manual verification and release.

Large collection range: The collection requires the car with the equipment to run on the road and pave the main road of the target market.

Data processing and manual verification: all need to be based on the situation in different regions, the process needs to be developed, the model is required and even manual labeling is required.

Data updates cannot be guaranteed:

After a large number of unmanned vehicles are put on the market, the latest data of the user car sensors through the road can be returned to ensure the freshness of the map.

The maintenance and iteration of subsequent manual/automatic updates of maps also incur significant costs.

2) The country is a large-scale, unified market, the marginal cost of mapping is low, and the update frequency is guaranteed

We can see that domestic manufacturers will choose the automatic driving solution of high-precision map.

The marginal cost of cartography is low: the specifications of the main arterial roads in the country are unified with "one sign, one marking line, and the whole country is common".

Domestic traffic density also ensures sufficient road network coverage, in reaching a certain penetration rate, sensor signals of different car manufacturers are summarized to the map provider / automatic driving operators, can achieve high-frequency road refresh.

The map industry involves state secrets, and the entry threshold of the fine map industry is very high

At present, there are fewer enterprises in China that have "Grade A surveying and mapping qualification for navigation electronic map production". However, it is precisely because of political barriers that foreign competitors are restricted from joining, which also gives domestic companies a certain advantage.

In order to avoid related problems, Tesla's "short-term road network" used for simulation training will be saved on domestic servers.

3) Compared with foreign countries, the domestic road environment is more complex and the requirements for safety are higher

Foreign land is sparsely populated, and the road environment is relatively simple;

Domestic road traffic is densely populated, with Traffic Participants with Chinese Characteristics (such as battery cars, pedestrians).

You may say that Uber, Waymo, Mobileye, and NVIDIA, which are also facing the global market, will use high-precision maps, and the market may not be the main reason.

Yes, Uber's taxi model means that the automatic control of vehicle scheduling only needs to meet the demands of a large number of users and major routes, Waymo's high-precision map base was developed from Google Maps, and Mobileye and Nvidia provide software and hardware for a large number of car manufacturers (perhaps meaning a lot of data recycling).

As an independent manufacturer, tesla, in addition to Musk's "first principle", may be unwilling to share the "soul" of autonomous driving (autonomous driving ability and user data) between each other, perhaps he has chosen a technical solution that can make up for the lack of high-precision maps.

Third, Tesla's pure visual solution: MIND OF CAR

If Tesla has and only has cameras, then his vision solution must be the highest level in the industry.

Tesla's train of thought is: people are driving cars, roads are designed for humans, and if they have the same ability to perceive and process information as humans, machines can seamlessly transition to autonomous driving, economics & advanced.

At Tesla's AIDAY 2021, Musk shared pure visual self-driving solutions and capabilities:

1) Hardware: 8 360 °, 120w pixel cameras, in the effect need to be able to replace the replacement of lidar, millimeter wave radar and sonar.

2) Perception: A labeling team of 1,000 people, 150w vehicles on the road Tesla, guarantees a large collection of data trained for visual algorithms.

Like most self-driving perception schemes, Tesla's vision scheme is doing the following things: image calibration Object recognition Depth perception creates a four-dimensional vector space.

However, Tesla chose to simulate radar through vision, and the accuracy rate is basically close to that of real radar.

Get the truth by running on the road with a radared road test car. After a lot of training, the algorithm can get the conversion relationship between vision and simulated radar, getting rid of the user car's dependence on radar.

Through vision, it is also possible to model unfamiliar roads, perceive and predict (visual algorithms based on a large amount of road data learn to obtain prediction information beyond the visual distance), and generate short-term high-precision maps.

3) Planning and control

Tesla is basically not much different from the industry on the planning rack, basically solving the feasible space first, and then using the optimization method to optimize the solution in the feasible space, and output the final trajectory.

Tesla's trajectory moving line planning can be simply understood as a search-based method to generate a large number of trajectories, comprehensive evaluation of safety, comfort, efficiency to select the optimal moving line, in the technical ability optimization can allow them to reduce the number of searches, improve efficiency.

It also enables the ability to learn the properties of different objects, distinguish between people/animals/vehicles, and predict their future direction through visual tracking, as well as derived prediction capabilities, making decisions and avoidance.

4) Computing power

Cloud Dojo super computing power: The existing Tesla car-side FSD chip computing power mainly relies on two SoC chips, the computing power is 144TOPS, which is not high, but the camera based on frame detection needs to rely on high computing power, and its core appeal is high bandwidth and low latency.

5) Simulation capability

This is the key to Tesla's efficient iteration of visual algorithms, generating real-time high-precision maps through vision, making virtual simulation maps, restoring the real environment, creating a virtual environment with more boundary scenes, and completing monster upgrading (neural network rendering) in the virtual world.

The map data returned by each Tesla on the road can become the original of the simulation map, which can be seen to be realistic to the construction effect of static and dynamic environments, and can be recreated and combined through AI capabilities. This provides a large number of boundary scenes and reduces the effort of labeling.

The more vehicles running on the road, the more data is collected, and the data in the simulation scene library will be exponentially multiplied.

四、At last

It is undeniable that Tesla has strong technical capabilities and has the industry's top level in algorithmic capabilities. However, Tesla's plan is a subset of the technical path of high-precision maps + multi-sensors selected by domestic manufacturers, and if the vision solution is mature enough, domestic manufacturers can also "turn left" at any time.

Means of transport do not have to make "legs" or wheels, and robots don't necessarily look like "people." A purely visual approach is not necessarily the most efficient approach. Even if autonomous driving eventually moves towards "driving like a human", in the early stage of the development of automatic driving and the process of "learning to walk", in the short term, with the help of high-precision maps and other sensors as "auxiliary walking braces", it is harmless to quickly achieve the safe popularization of automatic driving.

Resources:

"The First Unmanned Technology Book" by Liu Shaoshan

Blue Book of Autonomous Driving Simulation

Basic knowledge of autonomous driving series - HD map @ Shaolong https://>ata.alibaba-inc.com/articles/192451?spm=ata.23639746.0.0.3c0c4ddfCx53C1

Industry Frontiers: The "HIGH-Precision Map" of Autonomous Driving – PLUGANDPLAY Article – Know https://>zhuanlan.zhihu.com/p/486988101

High-precision maps are recognized as standard for autopilot, but Tesla doesn't care – https://>baijiahao.baidu.com/s?id=1685063016792310360&wfr=spider&for=pc

High-precision maps of autonomous driving (8) The development status of high-precision maps at home and abroad – Abao said the article of the car – zhihu https://>zhuanlan.zhihu.com/p/369259860

Tesla AI Day analysis with ultra-long delay: Explain the FSD vehicle-side perception – EatElephant's article – know the https://>zhuanlan.zhihu.com/p/458076977

Tesla AI Day Decision Planning Technology Analysis – Article on Autopilot Tractor – Knowing https://>zhuanlan.zhihu.com/p/402442178

How do you empower machines to think? Decrypt Tesla Artificial Intelligence Autopilot Next Episode! More than you can imagine https://www.>bilibili.com/video/BV1uU4y1u7JY/?spm_id_from=autoNext

This article was originally published by @Abai on Everyone is a Product Manager. Reproduction without permission is prohibited.

The title image is from Unsplash and is based on the CC0 protocol