GraspALL: Adaptive Structural Compensation from Illumination Variation for Robotic Garment Grasping in Any Low-Light Conditions

Haifeng Zhong; Wenshuo Han; Zhouyu Wang; Runyang Feng; Fan Tang; Tong-Yee Lee; Zipei Fan; Ruihai Wu

cs.ROMar 16, 2026

GraspALL: Adaptive Structural Compensation from Illumination Variation for Robotic Garment Grasping in Any Low-Light Conditions

Authors: Haifeng Zhong, Wenshuo Han, Zhouyu Wang, Runyang Feng, Fan Tang, Tong-Yee Lee, Zipei Fan, Ruihai Wu,

Organizations: School of Artificial Intelligence, Jilin University · Chinese Academy of Sciences · National Cheng-Kung University · Peking University ·

Abstract

Achieving accurate garment grasping under dynamically changing illumination is crucial for all-day operation of service robots.However, the reduced illumination in low-light scenes severely degrades garment structural features, leading to a significant drop in grasping robustness.Existing methods typically enhance RGB features by exploiting the illumination-invariant properties of non-RGB modalities, yet they overlook the varying dependence on non-RGB features under varying lighting conditions, which can introduce misaligned non-RGB cues and thereby weaken the model's adaptability to illumination changes when utilizing multimodal information.To address this problem, we propose GraspALL, an illumination-structure interactive compensation model.The innovation of GraspALL lies in encoding continuous illumination changes into quantitative references to guide adaptive feature fusion between RGB and non-RGB modalities according to varying lighting intensities, thereby generating illumination-consistent grasping representations.Experiments on the self-built garment grasping dataset demonstrate that GraspALL improves grasping accuracy by 32-44% over baselines under diverse illumination conditions.

Figures & tables

Figure 2: Structure maps generated by the Canny operator from identical scenes under different luminance levels.

Figure 3: Overview of our GraspALL. (A) describes the generation process of the Parametric Luminance Curve and luminance compensation features (in Sec. 3.1 ). (B) illustrates the generation process of structural compensation features adaptive to the input luminance (in Sec. 3.2 ). (C) outlines the generation process of grasp points based on the luminance and structural compensation features (in Sec. 3.3 ).

Figure 4: Flowchart of proposed depth-optimal search strategy.

Figure 5: Construction of our MIGG dataset. (a) shows the household scenes and eight categories of garment assets. (b) shows that we obtain images under different illumination conditions by randomly combining the assets with the scenes. We undergo degradation in multiple dimensions such as color difference ( $\Delta{\mathbf{E}_{00}}$ ), black crush, texture clarity, and color temperature, rather than simple luminance attenuation. The lower part of (b) describes the dataset’s degradation status and the number/frequency of each garment category. (c) presents the captured garment images, depth map and semantic mask map. “ Luminance ” is calculated by averaging the histograms.

Class	Luminance: 90 – 120				Luminance: 60 – 90				Luminance: 30 – 60				Luminance: 0 – 30
——	SegMiF	MRFS	AMDA	Ours	SegMiF	MRFS	AMDA	Ours	SegMiF	MRFS	AMDA	Ours	SegMiF	MRFS	AMDA	Ours
Glove	74.9%	77.9%	76.6%	84.4%	70.1%	71.1%	75.5%	84.2%	61.9%	62.5%	70.7%	83.1%	62.3%	61.8%	68.5%	81.7%
Hat	82.8%	80.1%	81.5%	86.4%	74.6%	78.7%	80.3%	85.8%	71.8%	70.6%	75.9%	85.5%	65.4%	68.2%	71.3%	84.7%
Scarf	76.0%	82.2%	79.1%	85.7%	73.3%	75.1%	77.0%	86.3%	65.4%	65.7%	73.1%	85.7%	66.3%	63.5%	68.6%	85.6%
Sock	70.9%	67.2%	72.8%	83.5%	66.2%	65.7%	68.0%	82.2%	62.3%	58.9%	65.5%	80.3%	56.9%	61.8%	59.3%	79.2%
Tie	72.5%	80.3%	79.9%	84.7%	70.7%	73.9%	74.4%	84.1%	64.1%	66.2%	71.1%	82.1%	60.8%	65.3%	66.9%	81.4%

Table 1: Quantitative comparison of semantic mask generation accuracy between our GraspALL and other baseline methods under different luminance levels. The best results are highlighted in bold , and performance improvements are shown in red .

Class	Luminance: 80 – 120					Luminance: 40 – 80					Luminance: 0 – 40
——	BiFCNet	SAM-M	ReKep	DarkSeg	Ours	BiFCNet	SAM-M	ReKep	DarkSeg	Ours	BiFCNet	SAM-M	ReKep	DarkSeg	Ours
Glove	9/15	10/15	8/15	11/15	14/15	8/15	9/15	8/15	10/15	14/15	5/15	5/15	6/15	7/15	12/15
Hat	11/15	10/15	9/15	13/15	15/15	9/15	8/15	8/15	9/15	13/15	6/15	7/15	6/15	8/15	13/15
Scarf	10/15	9/15	10/15	12/15	14/15	8/15	8/15	7/15	8/15	13/15	6/15	6/15	5/15	7/15	13/15
Sock	9/15	7/15	10/15	11/15	13/15	7/15	7/15	7/15	9/15	13/15	5/15	5/15	6/15	7/15	11/15
Tie	9/15	10/15	8/15	10/15	13/15	6/15	5/15	6/15	7/15	11/15	3/15	5/15	5/15	6/15	10/15

Table 2: Comparison of garment grasping accuracy between our GraspALL and other baseline methods under different luminance levels.

Figure 6: Comparison of semantic mask and grasping key points between our GraspALL and baselines under different luminance levels.

Model	PLC	LRL	SRL	$\mathcal{L}_{sc}$	$\mathcal{L}_{bce}$	mIoU	mGSR
Model-1	✗					65.4%	50.0%
Model-2		✗				71.3%	72.5%
Model-3			✗			68.5%	57.5%
Model-4				✗		64.9%	50.2%
Model-5					✗	68.7%	70.4%
Model-6	✔	✔	✔	✔	✔	82.6%	88.3%

Table 3: Ablation study of different components (Lum: 0-40).

Figure 7: Comparisons of Parameters, FPS, mIoU and FLOPs.

Figure 8: Differences in Grad-CAM at different luminance levels.

Figure 9: Setup and observation of real-world experiment.

Luminance	BiFCNet	SAM-M	ReKep	DarkSeg	Ours
Lu: 0 - 20	5 / 15	4 / 15	6 / 15	7 / 15	12 / 15
Lu: 20 - 40	7 / 15	5 / 15	7 / 15	10 / 15	13 / 15
Lu: 40 - 60	7 / 15	7 / 15	9 / 15	11 / 15	13 / 15
Lu: 60 - 80	9 / 15	8 / 15	10 / 15	11 / 15	14 / 15

Table 4: Real-world success rate under different luminance levels.

Figure 10: Explanation of using RGB in low-light conditions.

Method	Luminance $<$ 20	Luminance $<$ 40	#Params
$N=6,$ $R=256$	mIoU:80.8%	mIoU:81.5%	34.53
$N=12,R=256$	mIoU: 82.2%	mIoU: 83.1%	47.83
$N=18,R=256$	mIoU:82.5%	mIoU:83.7%	61.13
$R=128,N=12$	mIoU:80.3%	mIoU:80.9%	47.83
$R=256,N=12$	mIoU: 82.2%	mIoU: 83.1%	47.83
$R=512,N=12$	mIoU:81.4%	mIoU:82.0%	47.83

Table 5: The effect of different

N

and

R

on model performance.

Figure 11: Visualize the different luminance curves in the trained Parametric Luminance Curves.

Luminance	BiFCNet [ 48 ]	SAM-M [ 14 ]	ReKep [ 13 ]	DarkSeg [ 45 ]	Ours
Lu: 0 - 20	5 / 15	4 / 15	6 / 15	7 / 15	12 / 15
Lu: 20 - 40	7 / 15	5 / 15	7 / 15	10 / 15	13 / 15
Lu: 40 - 60	7 / 15	7 / 15	9 / 15	11 / 15	13 / 15
Lu: 60 - 80	9 / 15	8 / 15	10 / 15	11 / 15	14 / 15

Table 6: Real-world grasping success rate of different methods.

Luminance	SegMiF [ 16 ]	MRFS [ 37 ]	AMDA [ 46 ]	Ours
Lu: 0 - 20	51.2%	52.9%	55.1%	70.5%
Lu: 20 - 40	55.8%	58.3%	59.5%	71.3%
Lu: 40 - 60	59.4%	62.7%	63.8%	72.4%
Lu: 60 - 80	64.3%	65.6%	66.3%	74.3%

Table 7: Accuracy rate of mask generation by different methods.

	GraspALL ( mGSR )		DarkSeg ( mGSR )
Luminance	with ST	without ST	with ST	without ST
$0-20$	80.0%	66.7%	46.7%	33.3%
Luminance	GraspALL ( mIoU )		DarkSeg ( mIoU )
$0-20$	70.5%	61.4%	50.7%	36.9%

Table 8: Analysis for our transfer strategy (sim-to-real).

Figure 12: Garment images under different illumination conditions collected in real-world scenarios.

Figure 13: We use the method in [ 34 ] to transfer the model trained on the synthetic dataset to the real-world dataset. $\mathbf{Model}\left(\phi\right)$ denotes the model trained on synthetic data. $\mathbf{p}_{gt}$ represents the pseudo-labels generated by the model. The domain adaptation process is as follows: first, train the model using synthetic data; then, train the model using both synthetic and real data; finally, train the model using real data.

Method	Luminance	mIoU	mGSR
GraspALL	0 - 20	82.2%	82.5%
GraspALL + [ 30 ]	0 - 20	82.9%	83.3%
GraspALL + TSD-En	0 - 20	84.5%	86.6%
GraspALL	20 - 40	83.1%	84.1%
GraspALL + [ 30 ]	20 - 40	83.6%	85.0%
GraspALL + TSD-En	20 - 40	85.2%	87.5%

Table 9: The improvement brought by the TSD-En to GraspALL.

Figure 14: Visualization results with and without our depth map enhancement strategy (TSD-En).

Method	Luminance $<$ 20	Luminance $<$ 40	Average
GraspALL + [ 26 ]	17/25 (68%)	18/25 (72%)	70%
GraspALL + [ 5 ]	17/25 (68%)	19/25 (76%)	72%
GraspALL + [ 45 ]	19/25 (76%)	20/25 (80%)	78%
GraspALL + Ours	21/25 (84%)	22/25 (88%)	86%

Table 10: Grasping performance of different grasping strategies.

Luminance	DarkSeg	BiFCNet	GraspALL
$0-20$	59.2%	46.7%	70.5%
$20-40$	61.6%	54.5%	74.4%

Table 11: Analysis for our Grasping Strategy (in Simulation)

Figure 15: Comparison between our grasping strategy and mainstream strategies that take the geometric center as the grasp point.

Figure 16: Non-garment images (towels and shopping bags) captured under different illumination conditions in real-world scenarios.

Metrics	$\alpha$ =0.01	$\alpha$ =0.03	$\alpha$ =0.05	$\alpha$ =0.07	$\alpha$ =0.09
mIoU	78.2%	80.9%	82.8%	81.1%	79.6%
mGSR	78.3%	81.6%	84.2%	82.5%	81.6%

Table 12: The effect of different

\alpha

on model (Luminance:0-40).

Method	mIoU	Method	mGSR
SegMiF [ 16 ]	64.6%	BiFCNet [ 48 ]	52.4%
MRFS [ 37 ]	67.5%	SAM-M [ 14 ]	59.2%
AMDA [ 46 ]	68.9%	DarkSeg [ 45 ]	63.3%
GraspALL	86.7%	GraspALL	88.3%

Table 13: Performance of different methods on other objects.

Luminance	BiFCNet [ 48 ]	SAM-M [ 14 ]	DarkSeg [ 45 ]	Ours
Lu: 00 - 40	37.7% (4.8-7.5)	35.5% (3.4-5.5)	48.8% (5.3-8.7)	80.0% (2.9-3.6)
Lu: 40 - 80	44.4% (3.8-4.9)	40.0% (4.5-6.4)	53.3% (2.5-4.4)	84.4% (1.9-3.2)

Table 14: Statistical analysis of different methods. The red font indicates the accuracy fluctuation across three experimental rounds.

Figure 17: The test samples we collected for different types of image degradation (exposure, noise, and motion blur).

No Corruption			Mildly		Severely
mGSR	mIoU		mGSR	mIoU	mGSR	mIoU
80%	73.4%	Exposure	70%	69.7%	70%	64.3%
		Motion blur	60%	67.2%	40%	53.8%
		Noise	70%	70.8%	50%	49.5%

Table 15: Validation of GraspALL Robustness.

Robotic Grasping

Robotic grasping studies how robots perceive, plan, and execute stable, task-appropriate contacts with objects across varied shapes and conditions.

See all 311 papers

MMGNet: Multi-Modal 6-DoF Grasp Synthesis for Cluttered Scenes
Oct 26, 2026Rui Cao, Jia-Hui Pan, Yichuan Li +4Robotic Grasping Multimodal Sensor Fusion
Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Hong Kong, China · Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong, China · Sino-German College of Intelligent Manufacturing, Shenzhen Technology University, Shenzhen, China
Human Universal Grasping
Jun 15, 2026Kevin Yuanbo Wu, Tianxing Zhou, Isaac Tu +5HOI Generation Robotic Grasping
New York University · Tsinghua University · University of Michigan
Multimodal Adversarial Quality Policy for Safe Grasping
Mar 2, 2026Kunlin Xie, Chenghao Li, Haolan Zhang +1Safe Robot Manipulation Robotic Grasping
The authors are with the School of Information Science, Japan Advanced Institute of Science and Technology, Nomi, Ishikawa 923-1292, Japan

Oct 26, 2026J/K move · Enter open · S save

MMGNet: Multi-Modal 6-DoF Grasp Synthesis for Cluttered Scenes

Rui Cao, Jia-Hui Pan, Yichuan Li, Yunfan Yu+3

Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Hong Kong, China · Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong, China · Sino-German College of Intelligent Manufacturing, Shenzhen Technology University, Shenzhen, China

Robotic Grasping Multimodal Sensor Fusion

Open paper AskPDF arXiv

Grasping is a foundational capability in robotic manipulation. Scene-level 6-DoF grasp synthesis relies on point clouds for metric geometry, yet real-world point observations are often noisy, incomplete, and over-smoothed. RGB images provide object-region, boundary, and appearance cues, yet two bottlenecks limit their effective use in scene representation and local grasp reasoning. We present MMGNet, a Multi-Modal Grasp Synthesis Network that addresses these bottlenecks through hierarchical image-to-3D feature injection and rotational graspness modeling. Hierarchical injection integrates multi-scale image features throughout 3D representation formation, while the rotational graspness jointly predicts the grasp approach direction and in-plane rotation before organizing local image-informed features through grasp-shaded grouping. Ablation studies show that these two modules provide complementary gains. MMGNet achieves a mean AP of 60.34 on GraspNet-1Billion with RealSense, surpassing all compared methods. In real-world experiments with adversarial grasp objects, MMGNet improves the grasp success rate over AnyGrasp by 8.46 percentage points. Code, models, and data will be released upon publication.

Abstract

Robotic Grasping

Explore similar work

MMGNet: Multi-Modal 6-DoF Grasp Synthesis for Cluttered Scenes

Human Universal Grasping

Multimodal Adversarial Quality Policy for Safe Grasping