DeepMind who made alpha go releases AI "DeepMind Lab" playing 3D game as open source

Google acquired an artificial intelligence development company "Deep MindIs an artificial intelligence that broke down the Go champion "AlphaGo(Alpha Go) "has been developed for it. In order to contribute to further development of artificial intelligence, DeepMind is a flagship AI platform used inside DeepMind "Deep Mind Lab"Opened as an open source.

Open-sourcing Deep Mind Lab | Deep Mind
https://deepmind.com/blog/open-sourcing-deepmind-lab/

DeepMind Lab is an AI platform that learns by playing 3D games, and anyone can use it as open source. It is an "agent" that plays the 3D game given below, it has the body of the floating sphere, you can eject the thruster and move it in the direction you want to go forward and backward and leftward, and also take action like jumping and crouching. It is equipped with a camera of first person viewpoint, it is possible to change the viewpoint by moving the camera horizontally (right and left) and vertically (up and down). The agent learns based on surrounding images seen from the camera and on "compensation" and "punishment" obtained in the game, and makes decisions in real time.


As tasks of 3D games played, there are tasks such as "Collect fruits," "Go through the maze," "Promise dangerous pathways not to fall out of the cliff" "Jump around from the stage in space to the stage" "Let's mark with laser" to play games, "Learn and memorize randomly generated environments quickly" and so on.

In fact, you can see a sample movie playing a 3D game "DeepMind Lab" to get a "watermelon" goal.

DeepMind Lab - Stairway to Melon Level - YouTube


When the game starts, apples are lined up in front of the agent.


Looking back, lemon is also floating.


For the moment the agent seems to decide to move in the direction of the apple. When contacting an apple, the number of fruits you get will be counted in the lower right.


When you turn the corner there are several apples lined up ......


Go straight ahead and turn another corner, there is another apple. The end is a dead end ... ...


If you go to the front of the hotel and check it, it looks like it is not a height to reach with a jump, although it is a step.


After checking down a little, I found that there is watermelon on the step.


Then the agent returned to the way he came ... ...


I will move in the direction of lemon which was on the opposite side of the starting point.


As you turn around the corner, stairs are set up.


When I turned the corner that was ahead of the stairs, I was able to reach watermelon safely. In order to get the watermelon at the high place we found at the dead end, we decided to go back down the stairs and go up the stairs.


A movie shows how the agent of DeepMind Lab learns by playing a maze game with low difficulty.

DeepMind Lab - Nav Maze Level 1 - YouTube


The starting point is like this.


Properly going and there is an apple, we will get on while moving forward.


On the wall of the maze there are patterns to be markers ......


Red walls are installed.


I will go while getting an apple ......


There are also falling apples, and they are coming back and forth on the same road.


Eventually we discovered the unexplored route from the wall pattern, going forward ......


I found my first fruit. When we get this, the screen will be white out, and the agent learned what to do if we goal.


The second play started. There is a red wall in front of you.


This time I do not go to blindly, but a little turning from the start position, I found a yellow pattern near the goal in the first play.


There was a fruit which became the goal after bending the corner.


The third play started. In other words, this is seen as "a game starting the maze of the same shape from a different starting point". There is also a familiar red wall ......


I was able to goal quickly. By playing multiple times, you can see that the agent was learning well.


In the end, the movie is challenging a game with a high difficulty combining multiple elements "a laser tag game that jumps on a jump base on a foothold in outer space."

DeepMind Lab - Laser Tag Space Bounce Level (Hard) - YouTube


The start point is like this, space is spreading around, and a jumping stand in front of you. You can see multiple scaffolds around different heights.


When it reaches the jump table it fires (big jump) automatically, but the robot which was in the jumping position fired the beam. When you hits the beam, the middle number decreases ... ...


You can see an effect like an agent disappearing and over the game.


In the second play, there is an enemy robot in front of you from the beginning, and the beam is fired.


Successful collision avoided immediately toward the jumping base.


...... I think that two robots are waiting at the landing destination, and the game is over as it scrubs the beam.


The third time started from a safe point. First jump from the jumping base ......


I am checking the surroundings while jumping.


There was a robot at the landing destination and I got an attack.


When I dodged himself to escape, there were also two robots behind, and the game was over again.


As a result of repeated play repeatedly like this, in the end it was able to fly away from all the robots and fly to the upper footing.


In addition, it seems that the degree of difficulty is different in each of three sample movies, DeepMind Lab can highly customize the content that you want to learn. Customization can be done from GitHub, the URL of GitHub is coming soonThe page of DeepMind LabIt is due to be made public.