rose.blog

This one is really cool. At a high level the way it works is as follows. You have a camera trained on a human. The data taken by the camera is processed so as to detect features, which are generalized patterns of behavior. For example, a feature detector could be configured to detect the presence of anger in the human, for example by learning-based methods.

In addition, there is a robot, which is connected to the data processing system connected to the camera. This robot has effectors which control its actions. In this application, the effector controls are functions of the processed input from the camera, where the rules connecting the two are user-determined. This generic situation, where the robot’s behavior is conditioned upon the input from the feature detectors connected to the camera, maps to a constraint satisfaction problem as described here.

The way this would work is that the human / camera / robot system would generate optimization and satisfiability problems, to determine how the robot’s effectors should fire, and these problems can be remotely solved using Orion.

For example, you could acquire a Hansen Robotics Einstein, sit it him on your desk, train a camera on your face, use an anger feature detector that causes the Einstein robot to laugh harder the angrier you get.