The second installment of ABB Review’s “Buzzword demystifier” series looks at artificial intelligence. This guest article is by Derik Pridmore, CEO of Osaro, a San Francisco-based company partnering with ABB in the domain of machine learning.
I’m often asked “what is artificial intelligence?” and “why is artificial intelligence important for industrial robotics?” The term artificial intelligence seems to intrigue, confuse, and frustrate businesses that hope to use it to better serve their customers.
Before we define artificial intelligence, it is useful to define intelligence. Intelligence is the ability of an agent (human or otherwise) to achieve goals in a wide variety of environments.
For instance, when presented with a new goal or situation, humans are often able to achieve that goal. Their intelligence gives them robust capabilities. An artificial intelligence is an algorithm (computer code) that has this property – the ability to achieve goals in a variety of environments.
Adding to the confusion, we often hear the terms “machine learning” and “artificial intelligence” used interchangeably. They have become synonyms because learning is a very useful method for increasing intelligence. Machine learning simply means algorithms that are designed to learn from data. Two machine learning techniques which have shown particular promise in the past five years are deep learning and reinforcement learning.
Deep learning is a supervised learning technique that uses neural networks to form compressed and generalized representations of a data set. Reinforcement learning is a separate technique which takes a representation of an environment and a goal (called a reward function), and then uses trial and error to discover actions that maximize the reward. (It might also take in examples of how to achieve the goal – this imitation learning is how Osaro’s software works.) In theory, by combining these two techniques we can simultaneously learn a representation and discover how to achieve a goal.
The only questions that remain are how fast, how well, across what types of environments, and with what guarantees can we learn? This is the focus of most modern machine learning research.
With these definitions in hand we can understand why artificial intelligence is important to industrial robotics. If a business changes a part or workplace even slightly, today’s robots fail and need to be given a new set of instructions.
The algorithms controlling conventional robots have very little intelligence. Artificial intelligence will increase the range of tasks and environments in which robots can used. It will also decrease the total cost of robotic deployments by saving time and alleviating costly reprogramming.