Learning by Observation through System Identification

Abstract

In our previous works, we present a new method
\nto program mobile robots —“code identification by
\ndemonstration”— based on algorithmically transferring
\nhuman behaviours to robot control code using
\ntransparent mathematical functions. Our approach
\nhas three stages: i) first extracting the trajectory of the
\ndesired behaviour by observing the human, ii) making
\nthe robot follow the human trajectory blindly to
\nlog the robot’s own perception perceived along that
\ntrajectory, and finally iii) linking the robot’s perception
\nto the desired behaviour to obtain a generalised,
\nsensor-based model.
\nSo far we used an external, camera based motion
\ntracking system to log the trajectory of the human
\ndemonstrator during his initial demonstration of the
\ndesired motion. Because such tracking systems are
\ncomplicated to set up and expensive, we propose an alternative method to obtain trajectory information, using the robot’s own sensor perception.
\nIn this method, we train a mathematical polynomial using the NARMAX system identification methodology which maps the position of the “red jacket” worn by the demonstrator in the image captured by the robot’s camera, to the relative position of the demonstrator in the real world according to the robot.
\nWe demonstrate the viability of this approach by teaching a Scitos G5 mobile robot to achieve door traversal behaviour.