Project Goal: create a controller for balancing an inverted pendulum on a cart.

Githut repo here:

https://github.com/agillies8/pendulum_sim

CAD here:

Features include:

• pendulum simulated in omniverse
• pendulum visualized in rviz
• CAD from Onshape
• URDF exported from ros, also in a USD
• Angle of pendulum plotted vs time
• 5 types of controllers:
  • Human teleop
  • Linear classic - in progress -
  • Model based RL
  • RL based on state feedback
  • RL via camera observing the scene

Steps:

Step 1: Base setup:

• Install omniverse and ROS2. Then take some tutorials. This is a whole thing. More notes here:
• Create CAD of the pendulum
• Export the pendulum to a URDF
• Launch the URDF so it is available in Rviz
• Convert the URDF to a USD and make available in omniverse
• Control the pendulum base in omniverse from ROS2 teleop
• Receive feedback of the pendulum angle in ros2
• Have the URDF in Rviz reflect the state of the pendulum in omniverse

Step 2: Classic controller:

• Model the physical system and come up with control policy
• Side journey, calculating velocities for the state vector:
• have the pendulum self-balance, eg, troubleshooting the controller.
• Implementing the controller directly in Isaac Sim

Step 3: RL controller

For the RL controller, kicking it up a notch and using a double pendulum, since the single pendulum case is covered pretty thoroughly in a RL tutorial, and seems a super easy policy to learn.

Before starting, I created a new double pendulum URDF form Onshape, and getting it back into isaac sim.

• A few notes worth covering on that:

RL Approaches:

Approach using Isaac Lab

Approach of using cindy: