Waymo and automakers do not reach level 5?… Tuuling aiming for autonomous driving with a unique approach

delivery

The web camera is a mass retailer for several thousand yen.AI is a laptop

Autonomous driving has a rank division from level 1 to 5 set by SAE. In general, it is easy to imagine a development process that gradually increases its functions to the level 5 of fully automatic driving, but from the AI ​​research perspective, the level of thinking is not always the best. You may have heard that there are two types of self -driving approaches, for example, seeing related images. An approach that expands advanced safe driving support (ADAS) functions based on sensor technology and control technology, and traces the level of autonomous driving. It is an approach that aims for level 5 (fully automatic driving) at a stretch instead of such step -up development (mainly by AI technology). It is said that the former approach is to adopt a large number of automakers that do not go out of level 4 in the straightforward method, and it is said that Google and other IT companies such as Waymo are aiming for the latter automatic driving. These are the differences in the method, not which one is better. However, there is also an idea that the so -called level 5 autonomous driving cannot be realized in an approach that aims to gradually evolve ADAS technology and aim for fully automatic driving. According to the definition of SAE, level 4 is fully automatic driving under certain restriction conditions. Level 5 is a complete automatic driving without conditions. The difference is that if it is level 4, the route bus route, unmanned taxis specified for the running area, unmanned work vehicles in specific areas, such as construction sites, factory housing complexes and port areas, and levels, and levels. If it is 5, it is interpreted as an autonomous driving without restrictions on the driving area or geofense. Currently, most companies that are studying automatic driving and unmanned taxis are aiming to commercialize autonomous driving on the premise of high -precision 3D maps. If necessary, information such as inter -road communication and carrier communication is also assumed. This means that in the contrary, in places where 3D maps are not prepared, and in situations where cloud connections are not possible, autonomous driving cannot be performed (control withdrawal). It cannot come out of the level 4 frame of autonomous driving under restriction conditions. ◆ Japanese venture that dares to challenge a difficult approach: There are actually few companies that take an approach that aims for fully automatic driving from the beginning, regardless of the steps of the tuillering level. Waymo and Tesla are said to be close to this approach, but Waymo assumes map data. Tesla is a special approach that aims for "deemed Level 5" with an approach that expands the range of autonomous driving under level 2 restrictions. As will be described later, it is difficult to aim for level 5 at a stretch due to industry standards, safety standards, and laws and regulations. However, there is a venture company in Japan that boldly challenged this approach. A self -driving technology company called Turing. Turing is a company named after Alan Tuuling, the founder of the Charles Babedge (analytical institution) in the calculator. If Babbage is the inventor of the computing machine, Turing is a person who created the basic concept of a computer (Turing machine) that operates the current software. He is also a scientist who led the decoding team of German cryptocien Enigma under World War II. The name is also known for the "Turing Test" used for artificial intelligence tests and the "Tuuling Award", which is said to be the Nobel Prize of the calculator. Just as Tesla was named after Nicola Tesla, who once competed for hegemony in Edison and Electric Power Business (Power Technology), Turing is also an original company. The an autonomous driving car they work on controls only with cameras and laptops sold at electronics retailers. The company has just started and the system is a prototype, but the automatic driving they aim for is based on the camera. The reason is that "complete autonomous driving" is impossible in control based on information from 3D maps, clouds and road infrastructure. ◆ Leave the technology to leave it to AI to be cognitive and judged in detail in detail. The difference between the ADAS approach of autonomous driving and the AI ​​-based approach can be explained where the machine (system) is in charge of the operation operation. The operation operation is realized by three repetitions: "cognition", "judgment", and "action". In autonomous driving, the actuator only has an electronic and mechanical reaction (accelerator, brake, handle operation, etc.) by the control signal in any approach. ADAS -based approach automatic driving is applied to the cognitive portion, but the algorithm (program) is responsible for the judgment part. The aim of Turing is to use AI (image processing, route search, course decision, etc.) for cognitive and judgment. Turing CTO Shunsuke Aoki explains this difference as follows. "AIS cars controlled by AI, unmanned taxis, etc. are processed here, which processes cars and people by processing information from camera images, Lidar, and other sensors. It is only done. After that, deciding on a running route and performing accelerator and brake operation will be processed by the conventional ECU program. It is necessary to describe the rules for. It is called the rules -based algorithm, but we believe that this approach has a limit. Tiring approach is the camera image and sensor information: The decision to pass is also left to AI (machine learning / deep learning). "◆ Vehicles that can run autonomously with a notebook PC and web camera, that is, the conventional approach is aware of the situation and gives the computer to the computer. The computer operates the car under the conditions written in the program. The Turing approach determines the situation recognition and how AI moves the car. The former cannot deal with events that are not written in the program (stop processing and hand over control to humans by error or control). The latter can be performed according to the images and other information on the spot, even in situations where programs and map data are not. For example, in the rules base (conventional), if AI recognizes something moving, but it is not possible to identify people or cars, it will be in the rules. However, if AI by deep learning is processed until the judgment, it will be possible to step on the brakes and avoid collision courses from the overall situation. AI by deep learning is a multiple layer of many parameters at the same time, and outputs a certain value according to the combination of each parameter value. In the case of image recognition, it will make a judgment of dogs and cats depending on the output value (range of). In the Turing approach, the situation in the scene is comprehensively determined whether to step on the brake or turn the steering wheel. The reason is that their prototype cars can run unmanned (although only limited courses) with only cameras and laptops. Since the input information is only a camera, it can only be moved on a simple course, but in principle learning various situations will allow you to run on general roads. In the future, we would like to accumulate running experiments and learning on public roads, sophisticated sensor technology and vehicle control technology, and realize an autonomous driving car that exceeds Tesla. ◆ AI Shogi's brain is also an ambitious challenge, but Turing has excellent young engineers, including Aoki. Aoki was a graduate of Carnegie Melon University (CMU) and was involved in automatic driving development. CMU is a prestigious academic field of computer science and engineering. He also belongs to the National Institute of Informatics (NII) and Nagoya University. Aoki has a powerful partner. It is Kazunari Yamamoto who has the title of the co -representative CEO who played Turing with Aoki. Mr. Yamamoto is the developer of the AI ​​shogi "Ponanza" who broke the famous chess player. He is an AI company and a technical advisor to HEROZ, a listed on the First Section of the Tokyo Stock Exchange. The ADAS approach is an inevitable approach for a car manufacturer. They have a certain responsibility for consumer protection. In addition, in order to meet the standards of function safety and driving safety required by WP29 and ISO26262, it is necessary to become a step -based development method, including legal development. When considering the manufacturer liability and litigation, the manufacturer's autonomous driving is required to be reproducible to produce the same result with the same parameters. In the event of an accident, it is necessary to reproduce and verify that the system is not defect afterwards. Therefore, there are circumstances that a pure AI approach cannot be adopted. However, for this reason, the development of AI research and autonomous driving technology should not be restricted or regulated. It is not a science or engineering, which is the same as a certain school denies another school, which describes their initiatives as "play" by industry standards. Autonomous driving is a problem that takes time such as legal development and social acceptance, but if you look at the world, there are companies and groups that do autonomous driving and AI research with similar approaches. In order to improve economic development and technical skills, I would like to expand the soil where such ventures grow in Japan.

Response Shinji Nakao