Submit

Login Login

Register Register

Home / Journals / CMC / Online First / doi:10.32604/cmc.2026.080767
Journal Menu
Special Issues
Table of Content

All issues

Online First

2026

2025

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

2009

2008

2007

2006

2005

2004

Open Access

ARTICLE

LAH-Net: A Low-Light Aware Hybrid Network for Robotic Manipulation

Yingying Yu1,2,#,*, Jun Yuan3,#, Tong Liu1,2
1 Beijing Academy of Science and Technology, Beijing, China
2 Beijing Key Laboratory of Traffic Data Mining and Embodied Intelligence, Beijing, China
3 Independent Researcher, Beijing, China
* Corresponding Author: Yingying Yu. Email: email
# These authors contributed equally to this work

Computers, Materials & Continua https://doi.org/10.32604/cmc.2026.080767

Received 18 February 2026; Accepted 07 April 2026; Published online 29 April 2026

Abstract

Accurate grasp detection is fundamental to successful robotic manipulation. Existing methods achieve reliable performance under good light conditions. However, their performance in low-light environments suffers from severe degradation due to the diminishing discriminative ability of visual features. In this paper, a novel low-light aware hybrid network LAH-Net is proposed. It comprises an alternating transformer-CNN module (ATCM) between the encoder and decoder, and a knowledge distillation-guided low-light enhancement module (KDLEM) before the encoder, which is activated by an illumination gate under low-light conditions. To generate highly robust and synergistic features, the ATCM module facilitates the iterative exchange between the local representations from CNNs and the global contexts modeled by transformers. Additionally, a transformer-to-CNN adapter and a CNN-to-transformer adapter are designed for bidirectional feature alignment. Meanwhile, the KDLEM module employs a teacher-student framework where the simplified student network distills knowledge from the powerful teacher to enhance low-light adaptability and maintain computational efficiency. Moreover, we introduce a real-world low-light grasp detection dataset (RLGD) for algorithm evaluation, which contains over 70 objects captured under four distinct low-light conditions. Our method achieves 99.4% and 98.8% accuracy on the Cornell dataset and 96.05% accuracy on the Jacquard V2 dataset. It also attains an accuracy of 96.5% on the RLGD dataset and demonstrates strong generalization across various low-light intensity levels. The real-world experiments in low-light scenarios validate the effectiveness of the proposed method.

Keywords

Robotic manipulation; grasp detection; low-light enhancement; alternating transformer-CNN; grasping dataset

  • 59

    View

  • 10

    Download

  • 0

    Like

Related articles
Copyright© 2026 Tech Science Press
© 1997-2026 TSP (Henderson, USA) unless otherwise stated

Share Link