Suppose we consider industrial robots: the most complex task for a robot arm is to plan kinematics in a constrained space. Tasks can be like picking an object from a basket or bin, assembling asymmetric components one over the other, trying to enter a confined space, and carrying out sensitive tasks such as welding. Wait, what are we talking about? Aren't all of these already in place and the robots are actually doing a great job in these applications? That's totally right, but what happens is the robot is guided by a human worker to move to a certain position in such confined spaces. This is simply because it is, at times, intricate to mathematically model such environments and hence they are solved using robot kinematics. As a result, the robot arm would have a hard time trying to solve inverse kinematics and may have a chance of colliding with the environment. This may cause damage to both the environment and the workpiece.

This is where reinforcement learning can be helpful. The robot can initially take the help of human workers to understand the application and environment and learn its trajectory. Once the robot has learned, it can begin operating without any human intervention.

It's not just industrial robot arms: reinforcement learning can be used in different applications, such as in mobile robots to plan their navigation or in aerial robotics to help to stabilize drones with heavy payloads.

The next section will give us an introduction to MDP and the Bellman equation.