TU-TESL

• 22

General Motors' cruise announced in June that it would delay a taxi service plan after saying it would debut in 2019. But the setback did not stop more than 1000 of its employees from moving toward the goal of public offering. Hussein Mehanna, head of artificial intelligence and machine learning at cruise, said that in fact, work on the underlying systems and infrastructure had accelerated over the next few months. "We continue to show progress. Over the past six months, the passion and motivation I've seen shows that the people of cruise We want to provide the safest and most comfortable car ever, "said Mehanna, a senior Microsoft, Facebook and former snap engineering director. "The faster we update (these systems) in the driver's seat, the faster we can get to the point of releasing the car." To that end, the company detailed the work of its mapping team, which created high-definition maps to enable its cars to position themselves on the road. The maps discussed include two types of resources: three-dimensional tiling rendered from short and long-range lidar data and data encoded labels - they provide information about features such as lane boundaries, traffic lights (and their locations), and curb edges. It is this information that eases the processing burden on the car and enables it to focus on navigation. Cruise can also code the map according to "thousands" of interactions within a given area, so as to code the expected behavior of other vehicles. Its engineers use in-house production systems to test assumptions and iterate over different versions of features to assess their impact on vehicle performance, after which they scale new features across the map. These maps will soon be out of date, but for "complex" products and operational solutions. They detect real-world changes, such as construction and reconstruction projects, and within minutes send updates to every car in the cruise fleet. Cruise's other initiative, known as "continuous learning machine," aims to update the control policies of GM subsidiary cars (artificial intelligence "drivers") twice a week. The goal is to automatically detect the edge situations encountered during driving and embed these examples into the artificial intelligence driver so that its algorithm can improve itself. These visual algorithms also consider postures, from which they try to predict what people might do before or after crossing the road ahead. For example, they can detect people looking at their phones and infer that they are unlikely to cross right away or they may be distracted. They are even perceptual enough to see people dressed up as giant palms, Mehanna said, one of many unusual scenes Cruise's cars have encountered over the years while driving on public roads. "It's a negotiation process," he added. "We think that knowing [human] posture is extremely important for safe driving in cities like San Francisco, and I don't think people have to think too much about it." The latest news that a car in Uber killed a pedestrian a year ago highlights the importance of posture detection. According to documents released by the National Transportation Safety Board last week, Uber's autonomy does not have the ability to identify people walking outside the crosswalk. In a broad sense, Mehanna claims that in a real autonomous driving scenario, the latest environment is crucial. Cruise revealed earlier this year that it was testing computer vision and voice detection artificial intelligence to help its cars respond to passing emergency vehicles, one of several efforts aimed at enhancing vehicle situational awareness. One of the challenges of machine learning and its application is that autopilot is one of the most exciting and difficult problems of artificial intelligence on this planet. At the same time, we also hope that autonomous vehicles and their basic models can reach a very high level of performance, "Mehanna said. "Unlike most machine learning apps, these apps are either movie ads, or recommendations, or Amazon. If you show them a bad movie recommendation or organic search result, no one will be hurt. With self driving cars, the threshold is much higher. " Cruise continued to test cars in Scottsdale, Arizona and downtown Detroit, with most of the deployment concentrated in San Francisco. The company expanded rapidly, with the initial fleet of 30 driverless cars growing to about 130 by June 2017. Cruise has not publicly disclosed the exact number, but the company has 180 autopilot cars registered in California DMV. Three years ago, the documents obtained by IEEESpectrum showed that the company plans to deploy up to 300 test vehicles nationwide.