MSGID: 1:317/3 641bd5f1   
   PID: hpt/lnx 1.9.0-cur 2019-01-08   
   TID: hpt/lnx 1.9.0-cur 2019-01-08   
    Simulated terrible drivers cut the time and cost of AV testing by a   
   factor of one thousand    
    New virtual testing environment breaks the 'curse of rarity' for   
   autonomous vehicle emergency decision-making    
      
     Date:   
         March 22, 2023   
     Source:   
         University of Michigan   
     Summary:   
         The push toward truly autonomous vehicles has been hindered by the   
         cost and time associated with safety testing, but a new system   
         shows that artificial intelligence can reduce the testing miles   
         required by 99.99%.   
      
      
         Facebook Twitter Pinterest LinkedIN Email   
   FULL STORY   
   ==========================================================================   
   The push toward truly autonomous vehicles has been hindered by the cost   
   and time associated with safety testing, but a new system developed at   
   the University of Michigan shows that artificial intelligence can reduce   
   the testing miles required by 99.99%.   
      
      
   ==========================================================================   
   It could kick off a paradigm shift that enables manufacturers to more   
   quickly verify whether their autonomous vehicle technology can save   
   lives and reduce crashes. In a simulated environment, vehicles trained   
   by artificial intelligence perform perilous maneuvers, forcing the AV   
   to make decisions that confront drivers only rarely on the road but are   
   needed to better train the vehicles.   
      
   To repeatedly encounter those kinds of situations for data collection,   
   real world test vehicles need to drive for hundreds of millions to   
   hundreds of billions of miles.   
      
   "The safety critical events -- the accidents, or the near misses -- are   
   very rare in the real world, and often time AVs have difficulty handling   
   them," said Henry Liu, U-M professor of civil engineering and director   
   of both Mcity and the Center for Connected and Automated Transportation,   
   a regional transportation research center funded by the U.S. Department   
   of Transportation.   
      
   U-M researchers refer to the problem as the "curse of rarity," and they're   
   tackling it by learning from real-world traffic data that contains rare   
   safety- critical events. Testing conducted on test tracks mimicking   
   urban as well as highway driving showed that the AI-trained virtual   
   vehicles can accelerate the testing process by thousands of times. The   
   study appears on the cover of Nature.   
      
   "The AV test vehicles we're using are real, but we've created a mixed   
   reality testing environment. The background vehicles are virtual, which   
   allows us to train them to create challenging scenarios that only happen   
   rarely on the road," Liu said.   
      
   U-M's team used an approach to train the background vehicles that strips   
   away nonsafety-critical information from the driving data used in the   
   simulation.   
      
   Basically, it gets rid of the long spans when other drivers and   
   pedestrians behave in responsible, expected ways -- but preserves   
   dangerous moments that demand action, such as another driver running a   
   red light.   
      
   By using only safety-critical data to train the neural networks that   
   make maneuver decisions, test vehicles can encounter more of those rare   
   events in a shorter amount of time, making testing much cheaper.   
      
   "Dense reinforcement learning will unlock the potential of AI for   
   validating the intelligence of safety-critical autonomous systems such as   
   AVs, medical robotics and aerospace systems," said Shuo Feng, assistant   
   professor in the Department of Automation at Tsinghua University and   
   former assistant research scientist at the U-M Transportation Research   
   Institute.   
      
   "It also opens the door for accelerated training of safety-critical   
   autonomous systems by leveraging AI-based testing agents, which may create   
   a symbiotic relationship between testing and training, accelerating   
   both fields."  And it's clear that training, along with the time and   
   expense involved, is an impediment. An October Bloomberg article stated   
   that although robotaxi leader Waymo's vehicles had driven 20 million   
   miles over the previous decade, far more data was needed.   
      
   "That means," the author wrote, "its cars would have to drive an   
   additional 25 times their total before we'd be able to say, with even a   
   vague sense of certainty, that they cause fewer deaths than bus drivers."   
   Testing was conducted at Mcity's urban environment in Ann Arbor, as   
   well as the highway test track at the American Center for Mobility   
   in Ypsilanti.   
      
   Launched in 2015, Mcity, was the world's first purpose-built test   
   environment for connected and autonomous vehicles. With new support   
   from the National Science Foundation, outside researchers will soon be   
   able to run remote, mixed reality tests using both the simulation and   
   physical test track, similar to those reported in this study.   
      
   Real-world data sets that support Mcity simulations are collected from   
   smart intersections in Ann Arbor and Detroit, with more intersections   
   to be equipped.   
      
   Each intersection is fitted with privacy-preserving sensors to   
   capture and categorize each road user, identifying its speed and   
   direction. The research was funded by the Center for Connected and   
   Automated Transportation and the National Science Foundation.   
      
       * RELATED_TOPICS   
             o Matter_&_Energy   
                   # Automotive_and_Transportation # Vehicles #   
                   Transportation_Science # Engineering   
             o Computers_&_Math   
                   # Robotics # Computer_Modeling # Virtual_Reality #   
                   Communications   
       * RELATED_TERMS   
             o Computer_vision o Road-traffic_safety o Safety_engineering   
             o Scale_model o Artificial_intelligence o Wi-Fi o   
             Automobile_safety o Ethanol_fuel   
      
   ==========================================================================   
   Story Source: Materials provided by University_of_Michigan. Original   
   written by Jim Lynch.   
      
   Note: Content may be edited for style and length.   
      
      
   ==========================================================================   
   Journal Reference:   
      1. Shuo Feng, Haowei Sun, Xintao Yan, Haojie Zhu, Zhengxia Zou,   
      Shengyin   
         Shen, Henry X. Liu. Dense reinforcement learning for safety   
         validation of autonomous vehicles. Nature, 2023; 615 (7953):   
         620 DOI: 10.1038/s41586- 023-05732-2   
   ==========================================================================   
      
   Link to news story:   
   https://www.sciencedaily.com/releases/2023/03/230322140354.htm   
      
   --- up 1 year, 3 weeks, 2 days, 10 hours, 50 minutes   
    * Origin: -=> Castle Rock BBS <=- Now Husky HPT Powered! (1:317/3)   
   SEEN-BY: 15/0 106/201 114/705 123/120 153/7715 226/30 227/114 229/110   
   SEEN-BY: 229/111 112 113 307 317 400 426 428 470 664 700 292/854 298/25   
   SEEN-BY: 305/3 317/3 320/219 396/45   
   PATH: 317/3 229/426