linzhank / invpend_experiment

Inverted Pendulum Experiment

• Author: linZHank

Using gazebo to simulate inverted pendulum and control through ROS The URDF model of the inverted pendulum and gazebo model spawn were refered to this tutorial Thanks to Matthew Chan and Arthur Juliani for their brilliant tutorials.

Setup

My configuration was Ubuntu 16.04, ROS-Kinetic and Gazebo-7.0. Other combinations of Linux, ROS and Gazebo may work, but not guaranteed.

1. cd to the /src directory in ROS workspace (e.g. cd ~/ros_ws/src)
2. git clone https://github.com/linZHank/invpend_experiment.git
3. catkin_make, or carkin build if you were using Catkin Command Line Tools
4. source ~/ros_ws/devel/setup.bash
5. run roslaunch invpend_description invpend_rviz.launch to check the model in rviz; run roslaunch invpend_control load_invpend.launch to spawn the model in gazebo and initiate ros_control

Scripts

All python scripts is located at /your/path/to/ros_ws/src/invpend_experiment/invpend_control/scripts.

• cartpole_v0.py and cartpole_v1 configures the model to be a reinforcement learning ready enviroment.
• test_env.py is a good starting point to familiarize this environment. Currently, sending random command and sinusoidal command to the cart through velocity control are available/
• qtable_train.py runs Q-learning algorithm on the cart-pole, however parameters are undertuned at current stage.
• qtable_eval.py evaluates the Q table learned by qtable_train.py

Q-learning

Training Demo Video

Rewards Accumulation

Evaluation Demo