Improving world models for robot arm grasping with backward dynamics prediction

Yetian Yuan, Shuze Wang, Yunpeng Mei, Weipu Zhang, Jian Sun, Gang Wang*

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

1 Citation (Scopus)

Abstract

With the advent of Industry 4.0, intelligent manufacturing has emerged as a prominent trend for future development. The integration of intelligent manufacturing scenarios with reinforcement learning offers significant advantages and potential. However, in real world scenarios, reinforcement learning faces challenges in terms of sampling efficiency and potential mechanical damage. Model-based reinforcement learning demonstrates advantages that make it more applicable to real scenarios. In this study, we propose the improved Dreamer algorithm and develop a system for learning picking and placing operations using multimodal information. The enhanced algorithm efficiently mitigates the challenge of low sample efficiency and demonstrates remarkable efficacy in addressing high-dimensional state space problems. Furthermore, the policy acquired through reinforcement learning can be utilised for the manipulation of object with diverse geometries. To verify the effectiveness of the algorithm, we construct a production line simulation environment for robotic arm manipulation based on the Coppeliasim platform. Additionally, the advantages of the improved algorithm are further validated through the continuous complex control tasks in DeepMind.

Original languageEnglish
JournalInternational Journal of Machine Learning and Cybernetics
DOIs
Publication statusAccepted/In press - 2024

Keywords

  • Reinforcement learning
  • Robot arms
  • World model

Fingerprint

Dive into the research topics of 'Improving world models for robot arm grasping with backward dynamics prediction'. Together they form a unique fingerprint.

Cite this