Animals’ ability to learn makes them ideal models for studying animal communication as well as animal-robots interactions. Therefore, it is important to understand how animals communicate with one another and what cues they perceive to determine whether an interaction is mutual or not.
Like humans, animals also have innate behaviors that help them to survive in their natural environments. For instance, many different animals use alarm calls to announce the approach of a predator or other danger. Animals learn these alarm calls and use them accordingly in different situations.
Animal communication is a complex topic that research groups are only just beginning to understand, but there are already many examples of how this process works in non-human species. Researchers at MIT’s Computer Science and
What does communication in animals look like?
Let’s assume you are a predator in the wild. You are interested in finding out what makes the prey animals in your environment different from other prey, or what makes them different from other predators. One way to do this is to observe the prey and learn from the calls they make when they are alarmed. We can call this method “reflexive learning.”
For example, a predator may notice that the sound of a rattlesnake rattling or hiss creates a reaction in its prey. It may then imitate the sound to see what will happen. This is called “proactive learning” or “conditioning.” Predators use reflexive learning to figure out what prey animals can tell them about their environment, and proactive learning to learn more about the behavior of their prey. Predators that rely solely on proactive learning have to be good mimics and have to know what sounds in their environment indicate a threat. Also, they have to have good auditory systems that enable them to distinguish these sounds from other sounds in their environment.
How can we learn from animals to teach robots?
To create an animal-like system that can learn new things, we can use the same principle that allows animals to learn: reflexive learning. The first step is to use a model system where reflexive learning is already in action. For example, say you are trying to teach a parrot how to recognize a new object. You can’t just throw the object in front of the parrot and hope for the best. Instead, you have to teach the parrot the name of the object by repeating the name over and over again. In this way, the parrot learns the name of the object before it can even see it.
The next step is to try to mimic this process in the robot. The robot has to be able to pick up certain properties of the original object, like its shape or color, in order to make the robot resemble the original object. The robot may not be able to copy the object’s exact shape, but it can copy the general shape of the object. This shape information can then be used to generate a response that the robot thinks is coming from the original object.
A robot learns to talk through imitation and reinforcement
This is a really important and often-overlooked process that allows robots to learn to talk. When a robot is trained to associate a certain sound or symbol with a certain action, it associates the sound or symbol with that action. This association forms what’s called a “repertoire,” which is just a fancy way of saying “what you’ve been doing lately.”
A robot that has repeatedly performed an action will have a stronger association between the robot’s actions and a sound or symbol. When the robot hears the sound or sees the symbol, the robot associates them with the action that the robot performed, and therefore with the robot itself. This type of learning is called “instrumental” learning, because it’s based on how a robot can use things like sounds or symbols to tell the robot what action to perform.
Video recording workshop
You can try these out in our video recording workshop! Robots can be trained through direct observation, but can also be trained by providing reinforcement. Reinforcement means that the reward for performing a certain action comes from observing the action itself. In the video recording workshop, a robot performs a task based on reflexive learning.
A human then records the task by using a camera to show what the task is and how the robot performs it. The human can then play the video to the robot, triggering the robot’s reflexive learning. The robot can try out different actions until it finds one that it likes best. The robot can then be programmed to automatically perform that action when it sees the camera.
Conclusion
In this paper, researchers led by senior researchers at the Computer Science and Artificial Intelligence Laboratory (CSAIL) at MIT and Google AI, have described a way in which robots can learn to talk through imitation and reinforcement. The researchers described a model of communication in which a robot can learn to produce sounds based on a modality-agnostic imitation learning algorithm, combined with a reinforcement learning algorithm for forming associations between sounds and actions.
