Deploying a trained RL model to a real robot
After training and testing the robot in simulation, the final task is to deploy the trained model in the actual robot. There are different ways to do this. One of the sample references from the Isaac Lab website is shown in Figure 15.11. The robot should have a powerful computer, like an NVIDIA Jetson board, to interface sensors and run the model on it.
Figure 15.11: Reference diagram of interfacing an RL model to a real robot (source: https://isaac-sim.github.io/)
The important section of code we have to configure is a state estimator. This is the block delivering the list of observations used for training. Instead of using simulated sensor observations, we are providing real robot observations here. We can use ROS 2 drivers or Isaac ROS packages to get observations from robot sensors. Once the observations are extracted, they will pass into the model inference runtime, which generates the action. The commanded action will...
