From the course: Introduction to Artificial Intelligence

Robotics

From the course: Introduction to Artificial Intelligence

Robotics

“

- One of the best ways to connect with humans is to join us in the physical world. That's why robotics is one of the most interesting areas in artificial intelligence. Robotics is about having machines work on physical tasks. This can be lifting heavy objects in manufacturing or using robots that deliver food. Robots can even be vehicles, like self-driving cars or subway trains. Inventors have long been fascinated with finding ways to have machines behave like living objects. In the past, robots were just limited to highly specialized machines. They were used as welding machines in auto manufacturing. The auto plants near my hometown employed several specialized robots. Some could lift a car and install parts underneath, but none of them would ever be mistaken as intelligent. As impressive as they were, these robots were very limited in what they could accomplish. Unless they were programmed, they couldn't help a coworker open a car door or start painting the hood. They worked best for repetitive tasks. Robotics combined with machine learning gives us many more options. A machine can adapt to its environment and learn new tasks on the job. A basic example of this is self-driving vehicles. You couldn't program a car to react to everything it might see on the road. That's why the newest vehicles are using machine learning on an artificial neural network. These vehicles are outfitted with complex sensors that feed data into the network. It needs to understand all the different roads a vehicle might encounter. Then it needs to look at all the different people, animals, and other vehicles the car might find on these roads. Then the machine looks for patterns in successful driving. A car must react differently when a deer is crossing the road than when it sees a pedestrian walking a dog. That's why you often see self-driving cars with a human in the driver's seat. They supervise how the artificial neural network reacts to the data streaming in from the outside world. But like any new skill, it takes time for the machine to collect enough data. In artificial neural networks, this is often called training the network. Google famously said that they think of their self-driving cars not as a robotics problem, but as a data problem. It's true that to figure out how to get a car to steer left or right is simple compared to having a car understand when to turn left or right. Some robots don't need this level of complexity. That's why some of them just use good old-fashioned AI. Remember that this is an AI system that uses symbolic reasoning instead of machine learning. You would just try to program the robot to act intelligently. This is the difference between a Roomba, that's just programmed to avoid bumping into walls, and a self-driving car that really needs to understand the roads. Most robots today are still programmed like a Roomba and not learning like a self-driving car. That's because when you're in the physical world, there's a much higher price to pay if you make a mistake. So if you wanted to create a robot that distributes prescription medication, there's a huge cost to making errors. That's why many robots take the simpler approach and still benefit from symbolic systems and good old-fashioned AI.

Contents