Hands-On ROS for Robotics Programming

Self-driving cab example using the RL algorithm

As stated previously, we will explain the learning process using Q-learning because of its simplicity and physical sense. Bear in mind that the Q-learning algorithm lets the agent keep track of its rewards to learn the best action for every single state:

• Each time an action is taken from a given state, a reward is obtained according to P.
• The reward associated with each pair (state, action) creates a q-table that is a 500 x 6 matrix.

The q-value for a concrete pair state-action stands for the quality of that action in that state. Hence, for a completely trained model, we will have a 500 x 6 matrix, that is, 3,000 q-values:

• Every row represents a state.
• The maximum q-value in each row lets ...