Amazon's new technology can simulate thousands of deliveries a night, greatly speeding up robot training

When engineers train AI agents to play video games, they usually complete hours of human player play time in just a few minutes to stress test their performance. But when developing AI agents for the real world, this shortcut cannot be pressed. The new Scout food delivery robot that Amazon is testing is expected to break this situation and accelerate artificial intelligence training.

In an interview with The Verge, Scout Vice President Sean Scott explained that the company has created a detailed virtual map of the suburbs of the United States to accelerate the development of robots. The company has collected a lot of 3D data of roads, real road textures, and simulated the environment from sidewalks to storm drains.

“We can simulate thousands of food deliveries overnight without having to use and test the robot in a real-world environment,” Scott said. “The robot doesn’t actually know it’s in a simulation, it thinks it’s the real world, which is really cool.”

Scott said he was not aware of other companies trying this level of training, and that Amazon's internal training facilities include an indoor robot park and special equipment to test the flexibility of robot wheels. Being able to accelerate the training and development of AI agents is very important for Amazon, especially since Amazon entered the food delivery industry relatively late.

Amazon unveiled its six-wheeled food delivery robot in January, but a number of startups have been creating delivery robots in the field in recent years, which have been tested in small communities such as offices and campuses.