A Teacher-Student Markov Decision Process-based Framework for Online Correctional Learning

Cooperative System Identification via Correctional Learning

“…If no corrections are received while it follows the trajectories throughout the environment, the shifted objects did not actually have an impact on the features so the robot can correctly perform the tasks. If corrections are received but β, computed according to (7), is large, then the features are still correctly represented and the robot just has to update their importance theta for that task as in (8). If, on the other hand, β is small, no θ can explain the corrections and therefore the feature representation φ(R, o 1 , .…”

“…Learning from corrections is another way of learning from human input that can be a good complement and advantageous in many situations where real-time and taskspecific learning is needed [7], [8]. Methods to incorporate corrections in real-time to align robot and human preferences have been shown to improve performance and adaptability for HRI [2], [9], [10], [11].…”

“…So from the same correction u H , as before, the robot estimates the new u * H and β. Based on these it then updates the weights θ according to (8), and can resume its trajectory according to (3), now for Φ being the sum of the values of the aligned features φ( R, õ1 , . .…”