Reinforcement project


(Gil kijong) #1

Hi, :slight_smile:

This project is reinforcement learning project with turtlebot3.

This is not a finished project, but I would like to post the process of studying deep learning and applying it to robots.

I developed the source code with reference to the link below.

The most basic elements of reinforcement learning are action, state, and reward.

This is my environment.

  1. state : scan ranges data of LDS.
    action : [-0.8, -0.4, 0, 0.4, 0.8] of angular-z.
    reward : 0 of action is 10. another is -1 and if the turtlebot3 is within 15cm of the obstacle, it is -500.

  2. The turtlebot3 has always 0.5 value of linear-x.

  3. The turtlebot3 performs one random action of the specified actions.

  4. If the turtlebot3 is within 15cm of the obstacle, episode is done.

  5. Repeat training.

I expected the turtlebot3 not to hit obstacles and to drive in the map.

Actually, as the episodes repeated, the turtlebot3 seems to avoid obstacles. But, i think it is not perfect because I still do not have enough training.

And the interesting thing in process is that the turtlebot3 passed through obstacles at the end of the video(12:10).

I’ll post the project after modifying the environment and training.

If you have a advice or interest, please tell me.