Artificial Intelligence

2019-05-20

Beginner engineers in robotics are often try to program a direct policy. They want to tell the robot what to do in each situation. For example, if the task is to grasp a ball then the robot should open first the gripper, then move the hand to the ball and then close the gripper. This kind of direct policy isn't able to react to new situations. The more elaborate way in formalizing human knowledge are action models. For the example of a grasping task with the Baxter robot.[1]

The idea of tracking an existing action model is utilized by the “Learning from demonstration” community. In the first step, a human teacher creates the action model, and in the second step the given action model is monitored.[2] Tracking means, that the software can tell if everything is right with the events in the gant-tchart. The shown action can fit within the model or doesn't match the model.

References[]

  1. Lafleche, Jean-Francois, Shane Saunderson, and Goldie Nejat. "Robot Cooperative Behavior Learning Using Single-Shot Learning From Demonstration and Parallel Hidden Markov Models." IEEE Robotics and Automation Letters 4.2 (2019): 193-200.
  2. RSS Learning By Demonstration https://youtu.be/yj0BlDGIqe8