Researchers conducted ruffles to make Drone crash on 15 thousand times on purpose and teach how to fly to AI

"The fact that there is a big gap between simulation using drone and reality is because researchers are afraid of Drone's crash!" Researchers at Carnegie Mellon University (CMU) We are carrying out experiments that make drone automatically avoid obstacles by intentionally crashing 1500 times and letting the artificial intelligence machine learning of the recorded image data.

Learning to Fly by Crashing
(PDF file)https://arxiv.org/pdf/1704.05588.pdf

Drone Uses AI and 11,500 Crashes to Learn How to Fly - IEEE Spectrum
http://spectrum.ieee.org/automaton/robotics/drones/drone-uses-ai-and-11500-crashes-to-learn-how-to-fly

You can check the state of the drone, which you have learned mechanically with artificial intelligence and automatically avoid obstacles from the following movie.

Learning to Fly by Crashing - YouTube


Drone was placed in about 20 indoor environments and repeated 11,500 crashes in 40 hours.


The state of the flight was recorded with the camera installed in the drone, and the recorded image was later used for machine learning of the Convolutional Neural Network (CNN: Convolution Neural Network). Recorded video that shot the crash situation is divided into two parts that fly without problems and two parts just before causing a crash, depending on the situation, depending on the situation, "It is appropriate to go straight ahead" "Is it right to turn right?" Whether it is appropriate to turn to the right? "


In the movie, the flight of the drones after the learning is recorded. Drone flying a narrow passage.


On the way, you may hit Gonggong on the side wall, but change the direction of flight automatically, and advance without colliding head-on obstacles and walls.


Also found a person in the aisle.


When detecting a person as an obstacle, change direction quickly ......


With no harm to people, they are moving in a direction without obstacles.


Also, when you take a picture of a drone with a camera placed in a different place, it looks like this. Drone flying in a narrow passage.


I head towards the camera.


However, when I knew that the camera as an obstacle was in my direction of travel, I switched the course in the direction to advance without any obstacles.


Flying with Bu - no ... ....


I cleared the next aisle without difficulty.


I will skip the drone in the passage where the chair is located.


Proceed straight up to where there are obstacles ......


I will avoid avoiding chairs that are obstacles altogether.


Even if humans pass ......


I am flying with avoiding moving people.


As a result of the flight test, except for the complex environment of "placing multiple chairs in a narrow aisle", it was still impossible to fly as accurately as human manipulation. But,(PDF file)Monocular depth estimationIt is accurate compared with the situation, it seems that it showed performance 2 to 10 times depending on the situation. It is stated that it demonstrated excellent performance for places with glass doors and unspecified walls difficult to estimate depth.