Improving Reaching Tasks of a Simulated Fetch Robot Using Demonstrations along with Hindsight Experience Replay

Date

2019-05

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

Simple day-to-day activities like picking up or reaching out to an object seem easy for a human, but they are extremely difficult to teach to a robot. In order for the robot to do human-like activities with similar efficiency, the most popular option is Reinforcement Learning (RL). RL heavily relies on rewards to understand its surroundings. Most of the real-world tasks are naturally specified with sparse rewards and finding these rewards becomes extremely difficult as the task horizon and action dimensionality increases. These sparse rewards cause most of the RL algorithms to perform poorly. In order to achieve optimal performance while obtaining rewards, the proposed method utilizes demonstrations on top of Deep Deterministic Policy Gradients (DDPG) along with Hindsight Experience Replay, which substantially speeds up the training for simulated Robotics tasks.

Description

Keywords

Hindsight Experience replay, Deep learning

Citation