BCN3D-Moveo, Deep Learning

This repos aimed to enable deep learning for the open source-robotic arm BCN3D-Moveo.

Prerequisites

1. Installation of ROS Noetic

http://wiki.ros.org/noetic/Installation

2. Installation of Gazebo

http://gazebosim.org/tutorials?cat=install

Setup

1. Create a catkin workspace

http://wiki.ros.org/catkin/Tutorials/create_a_workspace

2. Clone this repository into the created workspace

git clone https://github.com/SLINIS/Moveo_RL.git

3. Build the project

catkin_make

4. Source workspace

cd catkin_ws
source devel/setup.bash

Start the Simulation

All launch files are placed in the moveo_moveit_config package. At the moment 3 different simulations for three different tasks exist, which can be launched by executing the commands provided below:

1. Inverse Kinematic

roslaunch moveo_moveit_config ik_training.launch

2. Push Task

roslaunch moveo_moveit_config push_training.launch

3. Pick Task

roslaunch moveo_moveit_config pick_training.launch

To start the training the following command needs to be executed:

python src/moveo_training/scripts/train_eval.py   --root_dir=~/sac/gym/MovoePush-v0/   --alsologtostderr

The flag --root_dir specifies the placement of the directory, where all data of the training and evaluation is placed. By using tensorboard the training and evaluation metrics can be plotted.

tensorboard --logdir ~/sac/gym/MovoePush-v0/ --port 2223 &