Thumb ticker md w01 small

Learning Robotic Manipulation from Instructional Videos

Vladimir Petrik (PostDoc)

The objective of the project is to enable robots to learn new manipulation skills from instructional videos available online. We will study how instructional videos could be used to overcome the sparse reward problem in reinforcement learning. The sparse reward complicates reinforcement learning by assigning the reward only after the task completion. Therefore, if the task is not completed successfully, there is no gradient guiding the reinforcement learning agent towards the goal. State-of-the-art approaches use human demonstrations to initialize the learning procedure. We would like to replace the human demonstrations with instructional videos that are already available, e.g. on Youtube, for the desired task. Although the task is the same, the key challenge is that the environment depicted in the video is not modeled in the simulation exactly. Instead, we will study how to modify the learning procedure to overcome differences in the simulator, real robotic testbed, and the video instructions.

Primary Host: Josef Sivic (INRIA and Czech Technical University)
Exchange Host: Ivan Laptev (INRIA, FR)
PostDoc Duration: 01 April 2019 - 31 March 2020
Exchange Duration: 01 October 2019 - 31 January 2020