oceanthunder/hoppip-v0 repository overview

⁠Hoppip-v0

This image provides a D4PG (Deep Distributed Distributional Deterministic Policy Gradient) RL training architecture for a one-legged hopper robot (Monoped-V0) in Gazebo/ROS(Robot Operating System), using Deepmind's Acme framework.

GitHub (Dockerfile): https://github.com/oceanthunder/Hoppip-v0⁠

⁠Setup

Pull the docker image:

docker pull oceanthunder/hoppip-v0:latest

Run the image:

xhost +local:root

docker run -it \
    --rm \
    --gpus all \
    --name hoppip-v0 \
	-p 6006:6006 \
    --privileged \
    --env=DISPLAY \
    --env=QT_X11_NO_MITSHM=1 \
    -v /tmp/.X11-unix:/tmp/.X11-unix \
    oceanthunder/hoppip-v0 \
    /bin/bash

Launch the Gazebo simulation:

export LD_LIBRARY_PATH=/opt/ros/noetic/lib:/opt/ros/noetic/lib/x86_64-linux-gnu
roslaunch my_legged_robots_sims main.launch

In a new terminal:

docker exec -it hoppip-v0 /bin/bash

Train using D4PG:

roslaunch my_hopper_training d4pg.launch

⁠Configs

If you want to use TD3/SAC/A2C, modify the start_training_v2.py file inside my_hopper_training/src and then run:

roslaunch my_hopper_training main.launch

If you want to tweak the rewards, modify the file at /root/monoped_ws/src/my_hopper_training/config/learn_params.yaml

Enjoy!