What are the risks in Uber's autopilot technology? Perceptual systems have always been unmanned

The specific cause of the Uber autopilot car accident is still under investigation, and the topic is continuing to ferment. Huang Chao, former head of the drip drone project and chief engineer and autographer and CEO of autowise.ai, analyzed the possible causes of the accident, from the difficulties of perception, system delay, and speed. Interpretation. "Blessings and blessings", in the case of Uber's rapid advancement, the whole industry is booming and promoting the commercialization of passenger cars, the emergence of accidents may allow major Internet companies, OEMs, Tier1 and other companies to calmly think about technical deficiencies. Sanitation and logistics commercial vehicles in parallel with passenger cars may be able to commercialize autonomous driving as quickly as possible.

The Uber hitting death on Sunday night continued to ferment.

Yesterday, the local Arizona director who was involved in the accident said that Uber may not be at fault in this accident, because from the video point of view, the parties directly traversed the lane from the dark, "it is difficult to avoid such collisions in any mode."

Although it seems that Uber is not at fault in this matter, Uber and the autopilot practitioners represented by Uber have experienced a collective public relations. The public is also worried: if the technology is not 100% perfect, Is there a collision accident in the road test? What are the risks in Uber's autopilot technology?

Yesterday, Huang Chao, the former head of drone drone project and chief engineer and autographer and CEO of autowise.ai, analyzed the high-risk "bug" that Uber and autopilot technology are prone to. And the deep-seated "big dry" problem behind the autopilot industry.

Three possibilities for Uber's car accident: perception is a problem, system delay, design speed is too high

1. Perceptual systems have always been the most difficult subject for driverless driving.

At the time, a 49-year-old woman planned to cross a two-way four-lane road, but she did not walk the crosswalk.

At this time, Uber modified Volvo XC90 SUV from south to north, the woman did not have time to dodge, and eventually a car accident.

The accident occurred on Sunday night and the surrounding sight was poor, which may be the main reason why the vehicle did not find pedestrians.

The autonomous vehicle that used the accident used a combination of 7 cameras and a high-beam laser radar. Huang Chao believes that this is a very typical sensor combination.

Uber's sensor system

Lidar and camera complement each other to a certain extent, and now the mainstream technical solution is also to integrate the two. According to the information obtained so far, the accident has a poor line of sight, so it may have an impact on the camera. At the same time, some unmanned systems will adopt some radical strategies when the speed is fast, thus reducing the sensitivity to the environment. degree.

The accumulation of these possibilities leads to the inaccurate detection or behavioral prediction of the pedestrian-side pedestrians in the entire perception system, which is even directly ignored, which in turn affects the decision-making layer.

2. The fatal delay in the appearance of the "brain" of the car receiving and processing signals

It is worth mentioning that after the sensor collects external data, it will send the data to the “brain” of the self-driving car, the central computing unit, and the “brain” creates a complete image around the car.

Huang Chao pointed out that the acquisition and transmission of data takes time. Taking the commonly used 64-line laser radar as an example, it takes 100 milliseconds to obtain each frame of data without optimization. Coupled with the time overhead of system calculations and the time spent on the underlying implementation of the car, the overall time delay is generally more than 100 milliseconds, or even seconds. And this delay is also related to the hardware state and the instantaneous load of the program. Therefore, it is not excluded that the accident is caused by the system entering a high-delay phase, and the sudden situation is not handled in time to cause a car accident.

3, the speed of self-driving vehicles is too high

Under US law, self-driving cars should also follow local speed limits.

Local police said that the Uber self-driving car reached 40 mph (about 64 km/h) at the time of the incident, while the speed limit of the incident street was 35 mph (about 56 km/h, but there were also reports of local With a speed limit of 45 mph, the 40 mph is still at a higher speed in urban suburban roads despite the low traffic at night.

In the first few years of Google, the test speed of unmanned vehicles was limited to 30 mph, because below this speed, unmanned vehicles are relatively safe even in the event of a traffic accident.

Huang Chao believes that at night, especially when some sensors are affected, the system should adopt a conservative strategy to slow down the car.

Moreover, Huang Chao emphasized that the unmanned system is a complete pipline from the perception to the execution layer, and the failure or abnormality of any link in the entire pipeline will lead to the failure of the entire system. At present, there is less information available, and it is not even possible to rule out hardware failures. Therefore, the follow-up police have a more complete investigation report, which can help us better analyze the cause of the accident.

In addition, the Volvo XC90 SUV is equipped with an automatic emergency braking system as standard, and it has certain obstacle recognition capabilities to prevent collision accidents. However, from the picture of the accident scene, the impact on the right side of the front of the vehicle was seriously damaged. Since there was no live video announcement, we were unable to determine if the car was braking urgently. Or when Uber is remodeling, this part of the original car has been greatly modified.

Behind the industry: the commercialization speed of passenger cars is slow, and the capital promotes automatic driving.

In the past few years, the domestic and international autonomous driving fields have ushered in the cusp of capital and technology. The enterprises represented by Uber can be described as driving all the way in automatic driving.

For the modified Volvo XC90 SUV, Uber has agreed to buy about 24,000 vehicles last year. The price of an XC90 SUV is about 50,000 US dollars. Buying these cars will cost nearly 1.2 billion US dollars. It is rare in the company of the car.

Recently, Uber also revealed the possibility of discussing with Toyota Motor about installing an automatic driving system in a Toyota.

Although Uber leads the number of autonomous vehicles, it rarely reveals the "quality" status, which is one aspect of Uber's widespread criticism after the car accident. According to the 2017 Autopilot Test report released by the California Highway Administration in February this year, Uber's opponent Waymo ranked first in the number of interventions per thousand miles, but Uber did not submit relevant data.

Domestic autopilots have also been hot under the blessing of capital and technology. According to incomplete statistics, more than ten autopilot startups have received financing last year.

Uber's car accident has also become a collective public crisis for autonomous driving practitioners. After Uber's car accident, nuTonomy, Toyota and other foreign research and development autopilot companies expressed their attitude to the test "speed reduction", and in response to the industry's "big dry" situation, Jingchi CEO Han Xu commented in the circle of friends:

We must have awe in doing automatic driving. I have always opposed the big deal. That's why I have been quarreling with experts in various coffee classes on various occasions: I have to insist on redundant multi-sensor fusion solutions. With the current computer vision technology, any madman who knows how to learn in depth can use the camera to do automatic driving at low cost.

On the one hand, Han Xu’s remarks show his concern about the squandering of individual companies and the excessive belief in the behavior of camera technology. On the other hand, it can also reflect that the “big dry” is behind the capital’s commercialization speed of passenger cars at this stage. The pursuit is speeding up. However, the real problem is that domestic autonomous driving is still in a high-speed development stage, and it will take a while for the final large-scale manned commercialization.

In addition to the passenger car, another wave of commercial vehicles, due to the limited environment and technical difficulties, the dawn of commercialization of autonomous driving.

In the market segments of commercial vehicles such as logistics, freight, and distribution, many players have been assembled, and many autopilot startups have also been born. For example, He Xiaofei, the former dean of the Institute, who recently left the company, founded the self-driving truck company's film-by-picture technology. There is still the main line technology founded by Zhang Tianlei and the future of Tucson; and the auto-driving logistics car and sanitation cleaning car. In the startup company, there was an autowise created by Chi Heng and Huang Chao.

Trillion market segment: The first autopilot sanitation field with less than 300,000 costs is commercialized?

Taking Huang Chao as an example, Huang Chao's autowise.ai focuses on the development of autonomous driving technology and uses the self-driving sweeper as the first commercial landing project.

As for why the auto-driving sweeper was chosen as the landing project, Huang Chao gave the following answers:

First of all, on the market side, the total length of roads nationwide has reached several million kilometers, with an average width of 10 meters, and the total area has reached tens of billions of square meters. The minimum price for ordinary road cleaning is usually 10 yuan per square meter per year. It is roughly estimated that the cost of road cleaning in the country is close to one trillion. At present, more than 60% of sanitation companies are labor costs.

Secondly, in terms of technology, the road cleaning route is relatively fixed, and the working time is mostly at night or in the early morning, and the road conditions are simple. The driving speed during cleaning is low, and strategies for protecting other traffic participants can be taken in the event of an emergency.

Finally, in terms of social value, road cleaning has always been the focus of urban sanitation work, but sanitation workers often need to work at night, even on holidays, and at the same time, because they are in an outdoor environment, they must not only face extreme weather, such as high temperatures. Cold weather, smog, and traffic accidents during work also occur from time to time. Automated driving sweepers can change this situation.

Autowise.ai's newly released driverless cleaning vehicle, including a 6-meter-long medium-sized cleaning vehicle and a 3-meter-long small cleaning vehicle, have begun trial operation in a large technology park.

The autopilot cleaning car in the video was converted from a nearly 6-meter pure electric sweeper and automatically started the cleaning task at two in the morning. During the operation, the cleaning vehicle can smoothly pass various traffic conditions such as traffic lights, roadside obstacles, etc., and automatically drive to the dumping dump to dump the garbage after the cleaning is completed, and finally return to the starting point to automatically park the parking space.

Self-driving cleaning car operation

Automatic parking

According to Huang Chao, in the future, most of the urban sanitation work will be completed by the self-driving clean car. When people wake up every morning, the whole city has already become a new look.

As for the feasibility of commercialization, Huang Chao said that at present, the total cost of all sensors for cleaning vehicles is only RMB 300,000. This is because the system uses multi-laser and low-cost GPS/IMU sensor fusion schemes for maps and positioning. Avoid the use of high-precision positioning equipment with hundreds of thousands of price points. At the same time, the sensing aspect uses a combination of low-beam LIDAR and camera fusion technology to replace the commonly used 64-line laser radar.

In addition, the current system uses a one cable sensor bracket solution to facilitate the transformation of a variety of models, and gradually move to mass production. Therefore, self-driving cleaning vehicles are most likely to be commercialized.

Based on this, the auto-driving cleaning team formed by autowise.ai has started trial operation in Shanghai.

Autopilot passenger car technology can be “down-division application” with more redundant hardware and software solutions to improve safety

Autowise.ai won the Red Dot China exclusive angel investment last year, and is currently conducting research and development of both self-driving passenger cars and commercial vehicles.

Huang Chao revealed to Xinzhiyuan that autowise.ai continues to develop unmanned passenger car systems because passenger cars need to handle more complicated road conditions and can collect more data, which can also accelerate the automatic driving of commercial vehicles. R&D, this is a process of “dimension reduction application”.

The autowise self-driving passenger car is able to travel smoothly on unstructured roads and complete the wrong car on the extreme roads, automatically passing through the toll gates.

At present, autowise.ai has more than 20 employees. The core strengths come from major Internet companies and car companies, including Google, Baidu, Didi, GM, Volvo, etc., in the fields of driverless, AI, travel network, big data, etc. Have a strong accumulation.

In addition to Huang Chao, the team also has the original Baidu computing team and data team architect Ye Qing as co-founder and chief architect. Ye Qing was responsible for Baidu's largest data warehouse construction and data mining project, and developed Baidu's first-generation large-scale Machine learning framework distributed vector computing engine.

From the passenger road test video given by autowise.ai, its self-driving passenger car has reached a considerable level, which can not only handle ordinary urban road conditions, but also handle some interesting unstructured road conditions.

Unstructured road car

Automatically pass the park toll gate

Huang Chao believes that the Uber event will promote continuous improvement of technology in the global autonomous driving field. At the same time, major Internet companies, OEMs, Tier1 and other companies will be calmly analyzed the current technical deficiencies, and more redundant hardware and software solutions will be used to improve the safety of autonomous driving technology.