- Initialize: Build a simulation environment with an inverted pendulum in gazebo
- Create inverted pendulum model using URDF, and visualize the model in rviz
- Spawn the inverted pendulum in gazebo
- Move the inverted pendulum using ROS
- Generate reinforcement learning friendly environment with ROS+Gazebo combination
- Read joint states from model
- Enables reset environemnt function (current major problem)
- Test velocity control on cart
- Feedback with reward for each state
- Implement deep reinforcement learning to control the inverted pendulum
- Q-table learning
- Q-network learning
- Deep Q Network
- Deep Deterministic Policy Gradient
- Build simulation environment with model of the robot in gazebo
- Create Kuka's model using URDF (partly available at kuka_experiment)
- Combine simulated robot model with real video data
- Use Kinect(or reasonable alternative) to record objects' moving trajectory.
- Simulated robot learns to catch with real human data feed
- Ultimate goal: The robot catches the ball thrown by human