FlatLab: A Unified Methodology Framework and Simulation-Based Benchmark for Robotic Manipulation of Flat Objects

Xingyu Zhu; Wenshuo Han; Zhouyu Wang; Yuran Wang; Ruihai Wu; Hao Dong; Fan Tang; Hechang Chen

cs.ROAug 14, 2026

FlatLab: A Unified Methodology Framework and Simulation-Based Benchmark for Robotic Manipulation of Flat Objects

Authors: Xingyu Zhu, Wenshuo Han, Zhouyu Wang, Yuran Wang, Ruihai Wu, Hao Dong, Fan Tang, Hechang Chen,

Organizations: Jilin University, Changchun, China · Engineering Research Center of Knowledge-Driven Human-Machine Intelligence, MoE, Changchun, China · Peking University, Beijing, China · Chinese Academy of Sciences, Beijing, China ·

Abstract

Robotic manipulation of flat objects is challenging due to the ungraspable configurations and strong variations in object geometry and material. Existing methods rely on heuristic pre-manipulation and are often evaluated in closed settings with limited generalization. We propose a unified framework that decouples the manipulation into a strategy generator and an action execution module. The strategy generator predicts appropriate manipulation strategies from object point clouds by learning strategy-centric, object-invariant representations via simulated data transformation and contrastive learning. Conditioned on the predicted strategy, the execution module decomposes long-horizon manipulation into reusable action primitives and dynamically composes them to generate stable trajectories. To enable systematic evaluation, we introduce FlatLab, a comprehensive simulation benchmark for robotic flat object manipulation. FlatLab provides high-fidelity physical simulation of diverse rigid and deformable flat objects, automated multi-modal data collection, and standardized task definitions and evaluation protocols. Experiments conducted in FlatLab demonstrate that our approach generalizes effectively to unseen objects and categories, outperforming existing baselines. The project page and the code are provided at https://flatlab-web.github.io/.

Figures & tables

Figure 1: Overview of FlatLab: a comprehensive simulation platform for robotic flat object manipulation. The platform offers high-fidelity physical simulation of over 100 flat objects and configurable manipulation scenarios. It includes our unified manipulation framework and supports automated multi-modal data collection, standardized task definitions, unified evaluation protocols, and turnkey deployment scripts. All functionalities are fully reconfigurable and extensible, facilitating ongoing research in flat object manipulation.

Figure 2: Motivation of Unified Flat Object Manipulation. Different flat objects exhibit distinct failure modes under single-strategy pipelines: edge-pushing fails for large or deformable objects, while dual-arm lifting is unreliable for thin or compliant ones. These challenges motivate a unified approach that adaptively selects strategies and executes actions for diverse flat objects.

Figure 3: Unified Robotic Flat Object Manipulation Framework. The Manipulation Strategy Generator takes point clouds of flat objects as input, embeds their features into a strategy-centric representation via multi-view consistency contrastive learning, and then predicts the manipulation strategy. The predicted strategy and the scene point cloud are fed to the Robot Action Execution Module. This module decomposes long-horizon manipulation into reusable action primitives, learns position-dependent poses rather than object-specific features, and then sequences the primitives to produce smooth trajectories.

Figure 4: Main Function Modules of FlatLab. First, FlatLab provides configurable assets with high-precision physical simulations, including flat objects, manipulation scenarios, and associated material textures. Second, FlatLab supports automated multi-modal data collection that yields perceptual and kinematic data from both successful and failed demonstrations. Third, FlatLab formalizes a set of robotic flat object manipulation tasks together with their evaluation criteria. FlatLab’s functionalities are fully reconfigurable and extensible, facilitating continued research.

Figure 5: Flat Object Dataset. The dataset is partitioned into three splits: a training set ( Train ), a test set with unseen objects from the same categories ( Test $\alpha$ ), and a test set with objects from unseen categories ( Test $\beta$ ). This partition enables a comprehensive evaluation of both object-level and category-level generalization.

Figure 6: Robotic Grasping Evaluation. Our method adaptively selects strategies to enable the grasping of flat objects of various shapes and materials, demonstrating significant generalizability.

Method	Train	Test $\alpha$	Test $\beta$
Slide	32.3%	27.3%	24.7%
Lift	41.2%	35.0%	27.2%
Diffusion Policy	48.8%	36.8%	38.8%
3D Diffusion Policy	66.2%	54.2%	51.2%
OpenVLA	55.5%	22.6%	17.8%
${\pi}_{0}$	63.0%	23.6%	17.3%

Table 1: Quantitative Comparison Results. Our method consistently outperforms all baselines on both training and test sets.

		A-1	A-2	A-3	A-4	A-5	A-6	A-7	A-8	Full
	RGB							$\checkmark$
Generator Input	$P_{obj}$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$			$\checkmark$
	$P_{env}$								$\checkmark$
Data Transformation		$\checkmark$			$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$
Contrast Consistency			$\checkmark$		$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$
Primitive Decomposition		$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$			$\checkmark$	$\checkmark$	$\checkmark$

Table 2: Ablation study results of the proposed unified robotic flat object grasping framework. The Manipulation Strategy Generator and the Robot Action Execution Module are evaluated separately through ablation. Strategy Accuracy denotes the strategy discrimination accuracy of the Manipulation Strategy Generator, while Grasping P-MSE denotes the prediction error of the grasp pose position in the Robot Action Execution Module. A-1 to A-8 correspond to different ablation configurations.

Figure 7: Real-World Experimental Scenario Setup. The experimental setup consists of a Baxter dual-arm robot, a table positioned directly in front of the robot, and an Intel RealSense D455 camera.

Figure 8: Real-world Evaluation of Our Approach. Experiments show that our method can successfully transfer to the real world, achieving high accuracy and strong generalization.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 9: Visualization of all flat object assets in FlatLab.

Figure 10: Visualization of various manipulation scenarios in FlatLab.

Method	Flat Object	Data Collection	Demonstration	m-Camera	m-Action	Policy
SAPIEN ( Xiang et al., 2020 )	$\times$	Manual	$\times$	$\checkmark$	$\checkmark$	$\times$
PlasticineLab ( Huang et al., 2021 )	$\times$	Manual	$\times$	$\times$	$\checkmark$	$\times$
FluidLab ( Xian et al., 2023 )	$\times$	Manual	$\times$	$\times$	$\checkmark$	$\times$
GarmentLab ( Lu et al., 2024 )	$\times$	Manual	$\times$	$\checkmark$	$\checkmark$	$\times$
ThinShellLab ( Wang et al., 2024b )	$\checkmark$ (D)	Automated	$\times$	$\times$	$\checkmark$	$\times$
GraspLargeFlat ( Wang and Kasaei, 2025 )	$\checkmark$ (R)	Automated	$\times$	$\times$	$\times$	$\checkmark$

Appendix

Table 3: Comparison of Simulation Platforms for Robotic Manipulation. Columns indicate support for flat objects (rigid R / deformable D), automated data collection, demonstration data, multiple camera views (m-Camera), multiple action primitives (m-Action), and policy execution. FlatLab is the only platform providing high-fidelity simulation of both rigid and deformable flat objects with automated multi-modal data collection, demonstrations, and full action-policy integration.

Figure 11: Dataset partitioning for all flat object assets in FlatLab.

	Train (S)	Train (G)	Test $\alpha$ (S)	Test $\alpha$ (G)		Test $\beta$ (S)	Test $\beta$ (G)
Book	20/20	19/20	18/20	16/20	Magazine	28/30	25/30
Disk	15/15	11/15	15/15	14/15	Pad	22/30	20/30
Plate	11/15	10/15	11/15	10/15	Disk Case	17/20	16/20
Box	16/20	16/20	15/20	14/20	Cutting Board	20/30	18/30
Board	15/15	13/15	12/15	10/15	Chessboard	15/20	13/20
Keyboard	15/15	12/15	12/15	9/15	Painting	18/20	14/20

Appendix

Table 4: Statistical analysis of experimental results on quantitative grasping of flat objects in a simulated environment. The results are recorded as “Number of successes / Total trials”. “(S)” indicates strategy discrimination, and “(G)” indicates grasping success.

Figure 12: Robot flat object manipulation procedures in major simulation environment scenarios.

Figure 13: Statistical results of the strategy discrimination success rate under different settings of the hyperparameters $\lambda$ and $\tau$ .

$\lambda_{rot}$ = 1.0	P-MSE	R-Err	P-MSE	R-Err	P-MSE	R-Err	P-MSE	R-Err	P-MSE	R-Err
Test $\alpha$	0.0048	0.0061	0.0047	0.0060	0.0038	0.0060	0.0034	0.0060	0.0040	0.0061
Test $\beta$	0.0058	0.0015	0.0045	0.0012	0.0041	0.0010	0.0040	0.0010	0.0043	0.0010
	$\lambda_{pos}$ = 1		$\lambda_{pos}$ = 3		$\lambda_{pos}$ = 5		$\lambda_{pos}$ = 10		$\lambda_{pos}$ = 20

Appendix

Table 5: Sensitivity analysis of pose position loss weights in robot manipulation. P-MSE measures the mean squared Euclidean distance between the predicted and ground truth object positions. R-Err measures the angular difference between the predicted and ground truth orientations.

		Slide	Lift	Diffusion Policy	3D Diffusion Policy	FlatLab (Ours)
	Book	19/20	10/20	14/20	13/20	19/20
	Disk	12/15	0/15	6/15	9/15	11/15
	Plate	8/15	0/15	4/15	7/15	10/15
	Box	0/20	18/20	10/20	14/20	16/20
	Board	1/15	13/15	12/15	8/15	13/15
Train	Keyboard	3/15	13/15	6/15	8/15	12/15

Appendix

Table 6: Comparison of quantitative experimental results for grasping evaluation in FlatLab across various baselines.

Figure 14: All real-world flat objects and their corresponding dataset partitioning.

	Train (S)	Train (G)	Test $\alpha$ (S)	Test $\alpha$ (G)		Test $\beta$ (S)	Test $\beta$ (G)
Book	5/5	5/5	5/5	5/5	Magazine	5/5	5/5
Disk	5/5	3/5	5/5	3/5	Pad	4/5	4/5
Plate	4/5	3/5	5/5	4/5	Disk Case	5/5	5/5
Box	5/5	5/5	4/5	4/5	Cutting Board	5/5	4/5
Board	5/5	5/5	5/5	4/5	Chessboard	5/5	3/5
Keyboard	5/5	3/5	5/5	4/5	Painting	2/5	2/5

Appendix

Table 7: Statistical analysis of experimental results for quantitative grasping of flat objects in real-world experiments. The results are recorded as “Number of successes / Total trials”. “(S)” indicates strategy discrimination, and “(G)” indicates grasping success.

Figure 15: Major real-world robot manipulation procedures.

Generalization in Robotic Manipulation

Generalization in robotic manipulation studies how robots transfer grasping and interaction skills to new objects, scenes, tasks, and embodiments.

See all 353 papers

AffordSim: A Scalable Data Generator and Benchmark for Affordance-Aware Robotic Manipulation
Apr 13, 2026Mingyang Li, Haofan Xu, Haowen Sun +10Affordance-Based Robotic Manipulation Robot Manipulation Datasets
School of Artificial Intelligence, Xi’an Jiaotong University
MobileManiBench: Simplifying Model Verification for Mobile Manipulation
Feb 5, 2026Wenbo Wang, Fangyun Wei, QiXiu Li +5Mobile Manipulation Robotics Simulation
University of Sydney · 1Microsoft Research Asia · 3Tsinghua University
ContactBench 1.0: A Benchmark for Everyday Contact-Rich Robotic Manipulation
Oct 17, 2026Jingyi Xiang, Yixuan Wang, Jeannette Bohg +1Contact-Rich Robotic Manipulation Robot Manipulation Benchmarks
University of Illinois Urbana-Champaign, Urbana, IL, USA · Stanford University, Palo Alto, CA, USA

Apr 13, 2026J/K move · Enter open · S save

AffordSim: A Scalable Data Generator and Benchmark for Affordance-Aware Robotic Manipulation

Mingyang Li, Haofan Xu, Haowen Sun, Xinzhe Chen+9

School of Artificial Intelligence, Xi’an Jiaotong University

Affordance-Based Robotic Manipulation Robot Manipulation Datasets Synthetic Data Generation for Robotics Robotic Manipulation Robot Manipulation Benchmarks Robotics Simulation+2

Open paper AskPDF arXiv

Many everyday robot manipulation skills are affordance-dependent, with success determined by whether the robot contacts the functional object region required by the subsequent action. Current simulation data generators obtain contacts from generic grasp estimators or per-object manual contact annotations, but generic estimators rank stable grasps without task semantics and often select contacts that are misaligned with the downstream action, while manual contact annotations must be rewritten for each new object and task. To solve these challenges, we introduce AffordSim, a scalable data generator and benchmark that integrates open-vocabulary 3D affordance prediction into simulation-based trajectory generation. Given a natural-language task description, AffordSim synthesizes a task-relevant scene, emits affordance queries, grounds them on object surfaces, samples region-conditioned grasps, and selects executable candidates with motion planning. It further randomizes object pose, texture, lighting, image noise, and cross-viewpoint backgrounds for sim-to-real transfer. We instantiate AffordSim as a 50-task benchmark across diverse manipulation skills, five robot embodiments, and 500+ rigid and articulated objects. AffordSim achieves 93% of the trajectory collection success rate of manual contact annotations on affordance-critical tasks and 89% on hard composite tasks. Vision-language-action policies trained on AffordSim data transfer zero-shot to a real Franka FR3, reaching 24% average success.

Abstract

Generalization in Robotic Manipulation

Explore similar work

AffordSim: A Scalable Data Generator and Benchmark for Affordance-Aware Robotic Manipulation

MobileManiBench: Simplifying Model Verification for Mobile Manipulation

ContactBench 1.0: A Benchmark for Everyday Contact-Rich Robotic Manipulation