WalkYTo-rl-gym

Walk Yourself, Toddler! Toddlers can learn by deep reinforcement learning now.

Setup Environment

Basic Environment

• ubuntu 18.04
• python 3.6

Including (core)packages

• gym-0.15.4
• roboschool-1.0.48
• tensorflow-2.2
• keras-2.4.3

Installation from Docker Image

you can get this environment by docker image, easily.

Pull image from the command line:

sudo docker pull docker.pkg.github.com/cun-bjy/walkyto-rl-gym/rl-gym:0.1

follow this page

How to Run Container

at the first time(make a container)

xhost +local:
sudo docker run -it --name gym --env="DISPLAY" --volume="/tmp/.X11-unix:/tmp/.X11-unix:rw" docker.pkg.github.com/cun-bjy/walkyto-rl-gym/rl-gym:0.1

• --env="DISPLAY" --volume="/tmp/.X11-unix:/tmp/.X11-unix:rw" : set to handle GUI environment from host, for Container.

after making container

xhost +local:
sudo docker start gym # if the container is on exit
sudo docker exec -it gym bash

Test Code on Container

#on the container
cd $HOME; python ant_v1.py

Trouble Shooting

• illegal instruction (core dumped) issue -> link

If you have some problem with this installation, issue up plz.

About Models

Model: RoboschoolAnt

similar to mujoco-ant in openai-gym

• observation

  • dim = 28
  • info =
    • position of Torso : (x, y, z)
    • orientation of Torso : (x, y, z, w)
    • joint angles : (hip 1,2,3,4, ankle 1,2,3,4)
    • linear velocity of Torso : (x, y, z)
    • angular velocity of Torso : (x, y, z)
    • joint velocities : (hip 1,2,3,4, ankle 1,2,3,4)

• actuators:

  • dim = 8
  • info =[hip_1, hip_2, hip_3, hip_4, ankle_1, ankle_2, ankle_3, ankle_4]

• rewards:

  • forward_reward : how far ANT goes on

  • control_cost : total actuator's toque

  • contact_cost : how many times the body contact on the ground

  • survive_reward : how long time the ANT goes on

    

What's New

2020-08-10

• walkyto-ddpg-keras submodule added in this package.

2020-07-04

• rl-gym:0.1 package released, this package w/ tensorflow-2.2 & keras-2.4.3

2020-06-20

• rl-gym environement package pre-released