TU-TESL

• 22

According to foreign media reports, a Silicon Valley start-up called ghost locomotion is quietly rising, joining the ranks of manufacturers and start-ups, dedicated to cars. However, unlike most companies in this emerging industry, Ghost Locomotion, founded in 2017, does not plan to introduce robot taxi services, nor does it intend to sell autopilot systems to suppliers and carmakers, nor is it intended to achieve automatic driving delivery. (photo source: Ghost locomotion) on November 7, local time, ghost locomotion announced that it had started from Keith rabois of founders fund venture capital, Vinod Khosla of Khosla ventures and Mike of Sutter Hill ventures private equity Speiser, which has received $63.7 million (about 444 million yuan) in investment, aims to retrofit mass-produced vehicles to be self driving. At present, ghost company is developing a suite that can let private car drivers drive automatically on the highway. The company said it would deliver the kit in 2020 and is expected to be compatible with 20 "popular" models after 2012. The price is not yet determined, but it will be lower than Tesla's automatic driving system (Autopilot), including FSD. Currently, the FSD price is $7000. Instead of equipping the car with an advanced driving aid system, the kit allows human drivers to give control of the car to a computer, and then the driver can play with his mobile phone or even doze off. According to founder John Hayes (former chief technology officer of Yahoo and co-founder of pure storage) and co-founder Volkmar Uhlig (Ph.D. in computer science, a senior employee of IBM Watson Research), the difference between ghost and other companies is that its automation method includes imitation learning. Unlike systems such as propilot assist and autopilot that rely heavily on rule sets, ghost first observes humans, records what drivers in the real world see, and how they react to create real situations. The company's artificial intelligence uses such information to build a model of correct driving behavior and create a dynamic automatic driving control strategy, which can be extended to almost any modern car. Ghost keeps a set of real-world data samples, which are not used to train the model, but to test, and is increasing more scenes, and retraining the model in order to achieve "perfection". Currently, the company's suite has been installed at the speed of dozens of commuters, LYFT and Uber drivers, and ghost plans to install the kit in hundreds of vehicles by the end of this year and thousands of vehicles by 2020. This method is similar to the method of wave, a Cambridge company in the UK, whose cars can improve themselves by learning the intervention measures of safe drivers (the system will learn from the mistakes of wave's system when the human driver takes over the system). Wayve said its system is far more scalable than companies such as waymo, Uber, cruise, zoox and Aurora, which are still using technologies that require a lot of data. For this reason, wave shows that cars on its platform can only use AI and satellite navigation to drive on the road. This is different from waymo's system, which relies in part on high-precision mapping and heuristics. (photo source: Ghost locomotion) ghost's upcoming kit will include a small computer installed in the car's trunk and connected to the controller area network (can), as well as eight low configuration cameras installed on the car's windshield, side window and rear window. This is no different from the Openpilot developed by Comma.ai. Like ghost's system, openpilot also runs on EON devkit and other third-party hardware, improving the car's computing power, enhancing driving aids such as sensors and keeping in the center of the lane and adaptive cruise control system. Similarly, X matik, a Toronto based company, is selling a similar aftermarket kit, lanecruise, which starts at $2000 and provides limited self driving capabilities for cars. The kit includes steering wheel controls, personal computers and cameras, and will automatically give up control once the driver delivers with the steering wheel or pedals. However, unlike openpilot and X matik's lanecruise, although ghost did not test its technology on public roads, the company said it tried its best to test its systems using processes commonly used in the aerospace and defense industries, comparing the driving model "numerically" with the way real drivers drive safely on public roads. It is speculated that such a comparison can help ensure that ghost will not violate the federal motor vehicle safety standard set by the national highway traffic safety administration. Three years ago, comma.ai was in trouble because it failed to comply with the standard. The whole method of ghost is based on the axiom that human drivers are basically right. First, a large amount of video data will be collected from the suite installed on the high mileage driver's car, and then the model will be used to find out what happened in the scene, and combine it with other data, including measuring the actions taken by the driver to determine the driving mode. Then, the model of ordinary driving is established, including the driving behavior of staying in the lane, braking and changing lanes on the highway, which does not need a long time or a lot of data. However, this does not "solve" the problem of automatic driving on the highway, because the most difficult part of automatic driving on the highway is how to build a driver who can handle emergencies (such as sudden turning) or correct such bad behaviors. If ghost can verify its system, it will be able to drive a private car automatically next year. Although NHTSA may step in. However, like Tesla, ghost's approach is not under the government's supervision, and the government has not yet regulated this area.