Tencent Technology News, June 9th news, according to foreign media reports, when engineers train artificial intelligence (AI) agents to play video games, they usually speed up, let the robot in a few minutes to complete humans can take hours to finish The game to test its abilities. Unfortunately, there is no such shortcut when developing AI that controls machines in the real world. You can't make reality "fast forward".
However, we can simulate reality, just as Amazon is doing to test its new delivery robot Scout. In an interview recently, Scout Vice President Sean · Sean Scott explained that the company has created detailed US suburban virtual maps to speed up the development of such robots. The company collected 3D real-time data and modeled the sidewalk into storm drains.
Scott said: "We can run thousands of analog deliveries overnight, rather than sending robots out to perform tasks in the real world. The robot doesn't actually know that it is in a simulated environment, but thinks that it is in the real world, which is a very cool progress. You can think of it as a ‘ hacker & rsquo; scene, but not just a delivery robot. ”
Scott added that he doesn't know which company "can achieve such high fidelity in this size of training", and Amazon's other training equipment also includes an indoor robot park and is used for testing. Special equipment for the flexibility of robot wheels.
For Amazon, it is important to speed up development, especially as it is a relatively late participant in the delivery robot game. The company did not launch six-wheel, cooler-sized robots until January of this year, but over the years, many start-ups have been manufacturing robots in this area and testing them in small communities and closed environments such as offices and campuses.
Of course, if a company can make this technology a reality, it is likely to be Amazon. Amazon founder and CEO Jeff · Bebes boasted that his company has the ability to invest huge sums of money for such projects, and Amazon likes to try strange delivery services, leaving the package at home. The driver went to the renovation delivery drone and so on.
But Scott said that the Scout project is focused on the opposite: ensuring that robots are not considered to be unusual machines, but part of the natural environment. He said that this was achieved to some extent through industrial design, and Scout's friendly curves, toy-like rubber tires and lively blue paint gave it a distinctly non-threatening impression. He said: "We hope that the robot disappears into the background, we call this ‘design for boring & rsquo;. ”
Figure 2: Another example of "design for boring" Amazon's AI Driver Training Park
This also means simulating the physical behavior of the Scout. Like all delivery robots, this machine must work with humans to navigate in ways that do not irritate or harm pedestrians. This means it will give way to children, pets, strollers and the elderly, and it will not make this respect indecisive.
This is a delicate balance that has been proven by self-driving cars. Studies have shown that pedestrians, drivers and cyclists may use any overly cautious software design to bully autonomous vehicles to cause them to leave the road.
Amazon is convinced that he is on the right path. For example, Scott proudly showed two videos of Scott encountering the same neighboring pet dog. Initially, the dog was curious about the robot and was wary of the outsiders in his territory. But then, it almost no longer pays attention to this robot. Scott said: "What better way to get a better response than pets?" ”
Figure 3: Some competitors (such as Kiwi Campus) have tested their robots for many years.
At present, Amazon's business is very limited compared to its competitors. It has only a few delivery robots working in Snohomish County, Washington, and every robot must be accompanied by a human supervisor to ensure that the robot does not get into trouble.
Amazon does not want to disclose details of the plan, such as where or when to extend these tests. But Scott said the company's simulations will help it enter new areas, enabling engineers to test the robot's autonomous driving skills in a virtual world before the robot hits an unfamiliar road. He said: "The next step is to capture other geographic areas [and] start the simulation test. Ultimately, we will be doing business around the world. ”
This is a hot topic, but it's clear that Amazon, like other companies in this space, still needs to solve many problems before shipping robots become common. In addition to issues such as reliability and security, there are some practical issues, such as how to send the package to the customer.
Amazon's Scout robots can't climb the steps, so the company's human assistants must now take the goods out of the robot and hand them over to the doorstep. Some companies are considering installing legs for robots to simplify the process. In addition, these delivery robots may also have regulatory challenges. Cities like San Francisco have labeled the delivery robots with “noisy” tags and completely banned them from driving on the sidewalk.
Scott said that the best way to prove that a delivery robot can work effectively is to make a delivery robot that can work in the virtual world first. He said: "We have a ladder to the moon, but we just boarded the first level of the ladder. ”“We have learned a lot, and we are just getting started. ”(Tencent Technology Review / Jinlu)