I recently read an article on robotics by a fellow intern. He cited an example of a flying bird robot developed by Festo Industries’ Bionic Learning Network. It is an autonomous ornithopter. While this is certainly an impressive achievement, because they have deciphered the flight of birds, something even more interesting has been going on in the field of machine learning. Developing robots is easy. What is difficult is making them learn how to do things on their own and improve their performance independently. That is what machine learning is all about.
Let’s take a look at an example here: Assume that you are designing the logic for a quadcopter. If you design the logic using the traditional approach, you will simply have to make sure that the quadcopter obeys the commands issued to it by the remote control. You would maybe go a step further and equip the quadcopter with onboard proximity sensors that would prevent it from crashing into the wall if the controller made a mistake. Though this might be good enough for running tests in your labs, it will not cut it in real life scenarios. Imagine that the quadcopter is being tested, and the battery is low for it to land safely. How would you ensure that it doesn’t crash? One way would be to have a sensor that would inform us about the battery status, another would be to have a reserve battery that would help us land the quadcopter if the primary one ran out. But this still leaves a large variable that has to be taken care of: human error. Time and again it has been proved that humans are really bad at performing tasks exactly as they’ve been instructed or repeating the task in the exact same manner. It turns out that computers are really good at doing these things.
Machine learning can be divided into two types: supervised learning and unsupervised learning. In supervised learning, we provide the algorithm with labelled datasets as references upon which its future actions may be built. For example, in the case of the quadcopter, we may define a set of flying patterns that consume less fuel. The algorithm then ‘learns’ the optimal flight path by analysing the given patterns. The unsupervised learning is entirely opposite of this; in this learning, we provide unlabelled data to the algorithm and it tries to make a sense out of it. For example, if we have a dataset of the temperature of the quadcopter body, taken while performing various stunts or while taking different flight paths, we can analyse this data and find out what exactly raises the temperature of the quadcopter and then take necessary steps to avoid it.
Besides quadcopter, this fusion of robotics and machine learning, also known as robot learning has found applications in a variety of fields, from defence to healthcare. In defence, robot learning can be used by military drones to formulate evasive maneuvers in case of an explosive situation. It can be utilized in a recon drone to sweep the area and suggest the safest path that can be taken by the troops. In the healthcare industry, robot learning could be used to teach internal healing procedures to nanobots because controlling a million different of these tiny robots is not humanly possible. In fact, we have already seen examples of such healing in popular culture such as Doctor Who, Ghost in the Shell: Stand Alone Complex, etc. We are bound to find many more applications of this remarkable technology in the future. To summarize, we are standing at the dawn of an entirely new era. And we have just begun to scratch the surface.
This article is part of PaGaLGuY’s innovative internship certification programme for engineering students. Currently, two such programmes are on – one is an Internship in Creative Writing, and the other is a Certification in Digital Media. If you are interested in partaking and bagging a certificate, besides learning the nuances of effective writing, write to us at [email protected]