Deep Mind releases a movie of AI figure that learns by myself only with the directive "go forward" and pushes with eerie movements

Google's DeepMind, which studies artificial intelligence,ReinforceWe announced the result of self-learning the technique to clear obstacles in AI using the paper "Emergence of Locomotion Behavior" in Rich Environments. A figure composed of rod-shaped parts, which clears the obstacles one after another with eerie movements brought up by themselves, is also released in the movie, making it a premonition of the evolution of the coming robot.

[1707.02286] Emergence of Locomotion Behavioral in Rich Environments
https://arxiv.org/abs/1707.02286

If you see the masterpiece movies that stick figures clear one after another, you can understand the magnitude of this technology with a single shot.

Emergence of Locomotion Behaviors in Rich Environments - YouTube


A model of only the body and feet made of movable rod shaped parts equipped with actuators.


I move my legs with dexterity and run, I get over the red obstacles well.


Although there is no eye, since it has the ability to sense the surrounding environment, you can grasp obstacles. For obstacles overhead (above the torso), we will lower the posture as if to go down, and we will move on and beyond.


The wall with the height also cleverly cleverly. It is an exquisite balance adjustment.


For high walls that can not get over well ... ...


Show her gesture like going backwards once, clear back and clear again.


I jump over a big gap and jump over it. The movement of this stick figure is not programmed. However, in order to maximize the rewards obtained from the environment by reading the surrounding environment, AI who was given the proposition "move forward", it is necessary to advance and move all parts by selecting and learning It is what became.


AI stick figures have learned how to succeed by changing the usage of the body by trial and error so that you can get more rewards while repeating failures such as falling, hitting a wall, falling into a rift, There are countless trials and errors repeating until I got ideal movement.


Just by giving an instruction "just move forward", AI learns how to move forward by reinforcement learning.


There are models like a spider with four spherical fuselages attached to the legs.


I will master the four legs with joints and will continue to jump repeatedly.


Of course, there is also a rift which can not be overcome because there is a physical limit in the movement of spiders and figures.


Spider figures can also overcome obstacles.


Human figure with 21 moving parts consisting of head, torso, arm and leg.


With the right hand sticking forward, the left hand moves forward with eerie move like pushing up on the diagonal upper side. Although it is how to use the body unfamiliar to humans, it seems that AI has gained this spooky movement as a way of running best for this course after repeating trial and error.


If you hit a wall ......


Human figure that falls. Even with troublesome obstacles, you can overcome any of them while training with reinforcement learning.


It seems like a human being, it looks like a summer like a hurd going beyond obstacles.


We will also incorporate hand-holding actions, and AI will find the best way.


It seems that a big gap is exceeded by a jump.


Of course there is a limit, but according to the "move forward" directive, AI will try to find a solution.


Human figure running round and round the left hand.



This AI learned various courses, but he said that he has never seen obstacles that move like a seesaw.


That the ground tilting and put weight, in spite of the obstacles for the first time experience, will continue to run in bold.


The red bars indicate the force applied from the outside.


While pushing or pulling, I will move forward while balancing.


For the walls that appeared in front of me ......


Try to bend your body and avoid it.



I will run up without stairs and stuff.


By giving diverse environments full of obstacles and setting a reward function according to progress, AI itself creates "behavior" that enables tasks to be executed efficiently This research will be effective in the future It is likely to be applied to robot technology that recognizes the situation and can better use the body.