How, exactly, do you teach a robot? Traditionally, to get a robot to do what you want requires programming– a list of tasks or rules to follow– which often means nuance and the unpredictable are left out. Or, in many AI systems, the robot learns from massive sets of training data. What if you could make a robot figure it out for itself, based on concepts and inference? Researchers have developed a new system for teaching robots concepts by presenting them a pair of images and having them figure out how to transform one to the other in real life. The goal is to introduce a method of learning used by humans all the time–learning by inference— like looking at the instructions for assembling an IKEA chair. Let’s take an easier example, these two images of apples – the input, on the left, and the output on the right.
For us, it’s easy to see how the apples are reorganized by position and color. The trick here is to make a robot recognize the conceptual differences, plan the changes needed, and then act on that plan. Researchers experimented with two different robots, each equipped with a camera at the end of a gripper for manipulating different objects on a tabletop. They presented the robots with a pair of images – an input and an output – representing a concept. Each concept consisted of different shapes, colors, and positions, which are captured by the camera and then run through special software.
It was then up to the robots to infer exactly how to rearrange the objects on the tabletop to match the output image in each concept. The robots achieved that by evaluating the different matchups in their imagination to arrive at a program that described the concept Researchers used a dataset of 546 different concepts to test the robots, and the robots successfully learned 535 of them. Researchers then tested 6 of these concepts with real-world variations in objects, backgrounds, and complexities. despite those variations, the robots were able to successfully execute the concepts They even found that they could teach the robots the steps needed by directing their attention with a pointer. Researchers believe that in the future, systems like these could be used to help robots achieve tasks without copying the movements of people and in new situations never encountered before…