| Title |
Deep Reinforcement Learning based Robotic Control via Object Positions Inferred from Raw Images |
| DOI |
https://doi.org/10.5573/ieie.2024.61.9.69 |
| Keywords |
Reinforcement learning; Position extractor; Convolutional neural network; Truncated quantile critics; Hindsight experience replay |
| Abstract |
This paper presents a method that utilizes raw images to infer the position of an object and uses reinforcement learning methods to move an object to a desired position by a robot arm. At first, for inferring the position of the object, pre-collected raw images and the exact position of the object are used to train a Convolutional Neural Network (CNN) Position Extractor, which is based on the feature extraction part of pre-trained VGG16. Then, reinforcement learning methods such as Truncated Quantile Critics (TQC) along with Hindsight Experience Replay (HER) are used to train a robot arm to move the object to a desired position in the simulation environment. A big advantage of this methodology is that an extra sensor for tracking the exact position of the object in real-time is not needed, since raw images are used to infer the position of the object and the robot arm uses that inferred position to move the object to the desired location. The mean cumulative rewards and the success rate of the methodology proposed by this paper are ?48.51 and 95.0% respectively. The results obtained by this method show that it is possible to move an object from a certain position to a desired one by a trained robot arm using the object positions inferred from raw images, instead of using the actual positions of the object. |