KinovaArm PegInsertTask with DRL

Train Kinova Gen3 arm to insert the plug into the board in Gazebo with deep reinforcement learning and imitation learning

Environmental preparation

We strongly recommend using docker to deploy environments, and we provide our environment zip pack docker.tar.gz.

• Get the docker environment

tar -zxvf dockerup.tar.gz
docker  load  <  dockerup.tar

• Get the graphical display package

sudo apt-get install x11-xserver-utils
xhost +

• Activate docker

docker run -d \
  -v /etc/localtime:/etc/localtime:ro \
  -v /tmp/.X11-unix:/tmp/.X11-unix \
  -e DISPLAY=unix$DISPLAY \
  -e GDK_SCALE \
  -e GDK_DPI_SCALE \
  --gpus=all \
  -e NVIDIA_DRIVER_CAPABILITIES=compute,utility \
  -e NVIDIA_VISIBLE_DEVICES=all \
  --name pytorch \
  dockerup
docker exec -ti pytorch /bin/bash

You can also deploy the environment step by step by referring to kinova

Train

• Go to the workspace directory

cd /catkin_workspace

• Start gazebo and load simulated arm

Load configuration file and open Gazebo to visualize， if you dont want visualization then add argument gazebo_gui:=false （Can't speed up training anyway).

source devel/setup.bash
roslaunch kortex_gazebo spawn_kortex_robot.launch gripper:=robotiq_2f_85 start_rviz:=false

• Start training

Open a new terminal and use the following commands to start training

source devel/setup.bash
roslaunch kortex_examples moveit_example.launch