YODO: Inverse Reinforcement Learning
Pramodith B, Tanmay B
Fall 2018 CS 8803 Interactive Robot Learning: Class Project
Georgia Tech
Our project aims to teach an agent to stack blocks together in a simulated 2D world. We attempt to make the algorithm partially extensible to new goal configurations. This can be abstracted as teaching an agent how and where to move objects located in any environment. A physical robot that operates in the real world would need to have skills to move objects around it to complete complex tasks. We developed a Reinforcement Learning Algorithm which is independent of the starting locations of each block in the environment to teach the bot to stack blocks. We then make use of an Oracle and expert demonstrations to help the Reinforcement Learning Algorithm based on the hypothesis that the oracle and human demonstrations can help in correcting the agent and thereby improve its performance.