Vision-Based Efficient Robotic Manipulation with a Dual-Streaming Compact Convolutional Transformer
Sensors · 2023
DOI: 10.3390/s23010515 ↗Kurz gefasst
Fachartikel in der Zeitschrift Sensors, 2023. Verfasst von Hao Guo und 4 weiteren Autor:innen. Beteiligte Einrichtung: School of Computer Science and Technology, Harbin Institute of Technology, Harbin 150001, China. Hauptthemen laut MEDLINE: Robotic Surgical Procedures, Robotics. Die Arbeit wurde bisher 5-mal zitiert (Crossref) und nennt 46 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Key Research and Development Program of China.
Abstract (Originalsprache)
Learning from visual observation for efficient robotic manipulation is a hitherto significant challenge in Reinforcement Learning (RL). Although the collocation of RL policies and convolution neural network (CNN) visual encoder achieves high efficiency and success rate, the method general performance for multi-tasks is still limited to the efficacy of the encoder. Meanwhile, the increasing cost of the encoder optimization for general performance could debilitate the efficiency advantage of the original policy. Building on the attention mechanism, we design a robotic manipulation method that significantly improves the policy general performance among multitasks with the lite Transformer based visual encoder, unsupervised learning, and data augmentation. The encoder of our method could achieve the performance of the original Transformer with much less data, ensuring efficiency in the training process and intensifying the general multi-task performances. Furthermore, we experimentally demonstrate that the master view outperforms the other alternative third-person views in the general robotic manipulation tasks when combining the third-person and egocentric views to assimilate global and local visual information. After extensively experimenting with the tasks from the OpenAI Gym Fetch environment, especially in the Push task, our method succeeds in 92% versus baselines that of 65%, 78% for the CNN encoder, 81% for the ViT encoder, and with fewer training steps.
Schlagwörter
MeSH-Schlagwörter (MEDLINE)
Stichwörter der Autor:innen
Autor:innen und Institutionen
- School of Computer Science and Technology, Harbin Institute of Technology, Harbin 150001, China.
- School of Medicine and Health, Harbin Institute of Technology, Harbin 150001, China.
ORCID-Profile: Hao Guo, Chunzhi Yi
Förderung
- National Key Research and Development Program of China
Zitieren
APA
Guo, H., Song, M., Ding, Z., Yi, C., & Jiang, F. (2023). Vision-Based Efficient Robotic Manipulation with a Dual-Streaming Compact Convolutional Transformer. Sensors. https://doi.org/10.3390/s23010515
Vancouver
Guo H, Song M, Ding Z, Yi C, Jiang F. Vision-Based Efficient Robotic Manipulation with a Dual-Streaming Compact Convolutional Transformer. Sensors. 2023. doi:10.3390/s23010515
BibTeX
@article{Guo2023,
author = {Guo, Hao and Song, Meichao and Ding, Zhen and Yi, Chunzhi and Jiang, Feng},
title = {Vision-Based Efficient Robotic Manipulation with a Dual-Streaming Compact Convolutional Transformer},
journal = {Sensors},
year = {2023},
doi = {10.3390/s23010515}
}
EUVIMED