On training intelligent agents in a virtual environment for the robot manipulator control task
Relevance. The necessity for developing effective methods for training intelligent agents in robotic manipulator control tasks in a virtual environment is critically important for enhancing the accuracy and efficiency of industrial and research processes across various fields. Aim. Implementation, i...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Tomsk Polytechnic University
2024-09-01
|
| Series: | Известия Томского политехнического университета: Промышленная кибернетика |
| Subjects: | |
| Online Access: | https://indcyb.ru/journal/article/view/61/50 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Relevance. The necessity for developing effective methods for training intelligent agents in robotic manipulator control tasks in a virtual environment is critically important for enhancing the accuracy and efficiency of industrial and research processes across various fields. Aim. Implementation, investigation and modification of reinforcement learning algorithms, such as Deep Q-Network and Proximal Policy Optimization, to manage agents in the KukaDiverseObjectEnv virtual environment on the PyBullet platform in order to create models that can accurately and reliably interact with objects of different classes. Methods. Programming, experimentation and synthesis, and comparative analysis. Results and conclusions. The authors have carried out comparative analysis of the effectiveness of Deep Q-Network and Proximal Policy Optimization algorithms, as well as modifications of the Proximal Policy Optimization for training agents in a particular virtual environment. It was shown that trained agents are able to solve the assigned task, and modifications can reduce the training time and the number steps required in the environment. As a result of testing, the algorithms demonstrated acceptable accuracy in manipulator control, which is validated on 1000 test episodes of the environment. The implemented algorithms and developed modifications have the potential to be used in industrial applications and further development in real-world conditions, which highlights their importance for modern robotics and automation. |
|---|---|
| ISSN: | 2949-5407 |