ddpgBaxter

This project aim at using the DDPG alogorithm in order to solve a simple graping problem with Baxter.

Setup info

*Ros kinetic
*Baxter simulator: http://sdk.rethinkrobotics.com/wiki/Simulator_Installation
*Keras 1.2.2
*Tensorflow
*CUDA 8 - CUDNN 6 for training

Rq : a contact sensor has been added to baxter's left gripper check the file in /src/baxter_common/rethink_ee_description/urdf/electric_gripper/fingers/extended_narrow.xacro

Done using :

*https://github.com/robosamir/ddpg-ros-keras
*https://robosamir.github.io/DDPG-on-a-Real-Robot/
*https://yanpanlau.github.io/2016/10/11/Torcs-Keras.html

The approach from https://robosamir.github.io/DDPG-on-a-Real-Robot/ has been adapted in order to control the speed of the robot and not the position

How to use:

Here the reaching task only has been solved. The concerned files are in the reaching folder of the ddpg package. (cf ddpg.py and new_robotGame.py)

*roslaunch baxter_sim_examples baxter_ddpg.launch
*rosrun ddpg ddpg

Rq in ddpd use play(0) for replay and play(1) for training

Project status

Reaching the cube ok