LLEGrasp: A Brain-Inspired Embodied Grasping Framework via LLM-Guided Frequency Disentanglement in Low-Light Scenes

Abstract

Grasp detection in low-light conditions poses great challenges for embodied agents due to weak visual information during complex physical interactions. While existing low-light grasp detection methods have achieved favorable performance in the pixel domain, they still suffer from blurred object edges and scattered distribution of illumination information due to their neglect of frequency-domain information, an essential component for enabling brain-inspired multimodal perception in embodied agents. To address these limitations, we propose LLEGrasp, an end-to-end brain-inspired framework that emulates the multi-scale filtering characteristics of the primary visual cortex (V1) to extract frequency-domain features under the cognitive guidance of LLM for low-light embodied robotic grasping. Specifically, it uses wavelet-based frequency separation to selectively harness boundary-enhancing noise patterns while suppressing destructive interference. Combined with illumination-aware bidirectional learning, it enables mutual refinement between raw and enhanced illumination features. Extensive experiments show that LLEGrasp yields respective improvements of 17.05% and 3.64%over the second-best methods under real-world and synthetic low-light datasets. Its effectiveness is also validated in real-world low-light robotic grasping.

Keywords

Low-light grasping / Grasp detection / Visual enhancement / Embodied intelligence

RIGHTS & PERMISSIONS

Higher Education Press 2026

Just Accepted

This article has successfully passed peer review and final editorial review, and will soon enter typesetting, proofreading and other publishing processes. The currently displayed version is the accepted final manuscript. The officially published version will be updated with format, DOI and citation information upon launch. We recommend that you pay attention to subsequent journal notifications and preferentially cite the officially published version. Thank you for your support and cooperation.