Panda Dagger

This repository implements a simple algorithm for imitation learning: DAGGER. In this example the agent is the Panda Emika robot and learn how to move in a goal position. The goal is to do a simple grasping task of a ring. It uses a monocular camera RGB to detect the ring position.

Getting started

This repositories need this fork of franka_ros repository, clone it and follow the standard install instruction. It is a modified version of the original repository with a robot model which works offline and a different kinematic solver (trac_ik_kinematics).

It has been used the robot simulator V-REP (version 3.5.0 proedu), download it here.

Optional but suggested step:

• Create a python virtual environment:

sudo apt-get install virtualenv
virtualenv venv
source venv/bin/activate

Nedded steps:

• Install python requirements:

pip install -r requirements.txt

• If you have Nvidia GPU with cuda you should install also:

pip install tensorflow-gpu

• Install ROS C++ dependencies:

rosdep init
catkin_make

Now you're ready to start!

Preliminary steps necessary for the tasks below

• Launch the controllers (this step is required for all node):

roslaunch panda_dagger panda_controllers.launch

• Open VREP (it is supposed you have an alias to vrep.sh)

vrep ./vrep/scene/panda_dagger.ttt

1. Generate dataset

• Run dataset generator:

rosrun panda_dagger dataset_generator

This node will iterate and will generate a dataset of optimal movement to reach the goal. You can stop the execution when you want, every single iteration is saved as single dataset in the dataset directory

2. Dataset merge

• Run dataset merge:

rosrun panda_dagger dataset_merge

This node will merge the datasets in the dataset folder into a single dataset called merged.pkl

3. Dagger

• Run dagger itself:

rosrun panda_dagger dagger

This node will execute the dagger algoritm. To customize execution look at the config file in config folder. When dagger execution is done it will save the trained net into the models folder.

4. Use trained model

• Execute trained model:

rosrum panda_dagger trained_controller