TU-TESL

• 22

According to foreign media reports, the L5 team of Lyft automotive research and development department revealed some details and progress of its automatic driving vehicle planning, and listed some challenges faced by its engineers. (image source: LYFT) imagine that the car was driving on the highway and met another car that crossed multiple lanes. If you want to prevent collision, you have to slow down, but how to control the speed of deceleration? LYFT's autopilot prototype uses a "human inspired" planning approach to determine deceleration rates. Initially, the company used a benchmark AI model that did not consider the speed of obstacles, while the latest model learned about human driving and gradually slowed down in the face of high-speed overtaking cars. LYFT level 5 team said, "this hybrid mode can be adjusted according to human preferences to make driving experience more comfortable and natural. We believe that combining rule-based systems, learning based systems, and human driving data can produce a comprehensive system level solution. " L5 engineers say that they are inspired by Maslow's hierarchy of needs theory and Asimov's three law of robotics, to design the system of automatic driving vehicle decision making. The pyramid model is based on safety and legitimacy, i.e. LYFT vehicles will verify whether their planned actions are safe and comply with local laws before performing operations. In addition, the concept of perceived safety is also considered in the planning model. Even when the actual safety risk is low, the insecurity of passengers and other drivers will be minimized. In actual driving, it may be necessary to widen the distance from the front car, or to ensure that the self driving vehicle will not be too close to the lane separation. The penultimate layer of the model is comfort, such as reducing nausea caused by gravity. The top of the model is route efficiency, such as arriving at the destination as soon as possible. The transparency of the automatic driving system will be the key to improving public acceptance. According to a survey commissioned by PSB Research last year, although about 94% of the accidents were caused by human errors, only 21% of Americans were willing to replace their cars with autopilot cars. Nearly half (43%) felt unsafe near the self driving car. The Lyft Level 5 team is made up of data scientists, application researchers, product managers, operational managers and other personnel. It is dedicated to building autopilot systems for carpool services. Since its establishment in July 2017, the Department has developed a new 3D segmentation framework, a new method for evaluating vehicle energy efficiency, and a technology for tracking vehicle movements using crowdsourcing maps. Earlier this year, LYFT opened a new road test site near its level 5 headquarters in Palo Alto, California. Engineers will simulate real driving scenarios in the center, including intersections, traffic lights, road merging, pedestrian access and other public road conditions, and the components will be resettable. A year ago, LYFT expanded its range of employee automated driving services in Palo Alto, with autonomous vehicles equipped with human safety drivers operating in restricted areas. This year, the company said it has tripled the number of routes available and plans to rapidly expand the coverage of the service. In November, Lyft revealed that its mileage increased by 4 times compared with half a year ago, and the number of employees engaged in automatic driving technology development increased from 300 to 400. According to the company, 96% of those who tried to use autopilot through Lyft application indicated they wanted to use it again. In May of this year, Lyft worked with Waymo, and Phoenix users could use Lyft to call Waymo's autopilot. In addition, Lyft is also working with Ann Andrew, who provides a small batch of self driving cars for customers in Lyft Las Vegas. Recently, Lyft has also opened its autopilot vehicle data set nuScenes, which has more than 55000 manually tagged 3D annotation traffic main frameworks.