# learning_efficient_robotic_garment_manipulation_with_standardization__20e588d3.pdf Learning Efficient Robotic Garment Manipulation with Standardization Changshi Zhou 1 2 3 Feng Luan 1 2 3 Jiarui Hu 2 3 4 Shaoqiang Meng 1 2 3 Zhipeng Wang 2 3 4 Yanchao Dong 2 3 4 Yanmin Zhou 2 3 4 Bin He 2 3 4 Garment manipulation is a significant challenge for robots due to the complex dynamics and potential self-occlusion of garments. Most existing methods of efficient garment unfolding overlook the crucial role of standardization of flattened garments, which could significantly simplify downstream tasks like folding, ironing, and packing. This paper presents APS-Net, a novel approach to garment manipulation that combines unfolding and standardization in a unified framework. APS-Net employs a dual-arm, multi-primitive policy with dynamic fling to quickly unfold crumpled garments and pick-and-place(p&p) for precise alignment. The purpose of garment standardization during unfolding involves not only maximizing surface coverage but also aligning the garment s shape and orientation to predefined requirements. To guide effective robot learning, we introduce a novel factorized reward function for standardization, which incorporates garment coverage (Cov), keypoint distance (KD), and intersectionover-union (Io U) metrics. Additionally, we introduce a spatial action mask and an Action Optimized Module to improve unfolding efficiency by selecting actions and operation points effectively. In simulation, APS-Net outperforms state-of-theart methods for long sleeves, achieving 3.9% better coverage, 5.2% higher Io U, and a 0.14 decrease in KD (7.09% relative reduction). Realworld folding tasks further demonstrate that standardization simplifies the folding process. Project page: https://hellohaia.github.io/APS/. 1Shanghai Research Institute for Intelligent Autonomous Systems, Tongji University, Shanghai, China 2The National Key Laboratory of Autonomous Intelligent Unmanned Systems, Tongji University, Shanghai 201210, China 3The Frontiers Science Center for Intelligent Autonomous Systems, Shanghai 201210, China 4College of Electronics and Information Engineering, Tongji University, Shanghai 201804, China. Correspondence to: Yanmin Zhou . Proceedings of the 42 nd International Conference on Machine Learning, Vancouver, Canada. PMLR 267, 2025. Copyright 2025 by the author(s). 1. Introduction Garment manipulation(Zhang & Demiris, 2022; Hoque et al., 2022) presents a critical challenge in robotics, with applications ranging from household assistance(Xiao et al.) and healthcare(Moglia et al., 2024; Wang et al., 2025) to industrial automation(Arents & Greitans, 2022). Unlike rigid objects(Yang et al., 2024a), garments are deformable(Jing et al., 2023), exhibit complex dynamics(Lin et al., 2022; Zhou et al., 2025b), and are prone to self-occlusion(Huang et al., 2022), which makes them particularly difficult to manipulate. These challenges are amplified when tasks such as folding(Xue et al., 2023), ironing(Canberk et al., 2023), sorting(Jing et al., 2024), and packing(Chen et al., 2024) require precise control over the garment s shape, orientation, and coverage. While significant progress has been made in garment unfolding(Zhu et al., 2023), existing methods especially those based on single-arm systems(Salhotra et al., 2023; Chen et al., 2023; Yang et al., 2024b) often require many interactions and long execution times. Dual-arm systems(Ha & Song, 2022; He et al., 2024) that leverage dynamic fling actions have been introduced to speed up the unfolding process, but they focus primarily on maximizing surface coverage, neglecting the crucial task of garment standardization. Standardization aligning the garment s shape, ensuring consistent orientation(Gu et al., 2024b), and achieving maximum surface coverage is essential for successful downstream tasks, such as folding and ironing. Without standardization, further manipulation becomes increasingly difficult, leading to suboptimal outcomes. To address these challenges, we propose a novel approach for garment manipulation, called the Action-Primitive Selector Network (APS-Net), which learns to predict spatial action maps actions defined on a pixel grid using overhead RGB-D images. APS-Net integrates both garment unfolding and standardization into a unified framework. By employing a dual-arm, multi-primitive policy, APS-Net intelligently selects between two key actions: dynamic fling and p&p (see Figure 1). The dynamic fling action is used to rapidly unfold crumpled garments, while p&p ensures finetuned alignment and standardization, preparing the garment for downstream tasks such as folding. Learning Efficient Robotic Garment Manipulation with Standardization Figure 1. Action Primitives. The fling action involves the robot grasping two predefined points on the garment, lifting and stretching it, then performing a controlled fling motion forward and backward while gradually lowering it onto the workspace. The p&p action involves the robot grasping the garment at a specified pick point, lifting it, moving it to a target place point, and releasing it. In the fold action, keypoint detection selects the pick and place points, and both arms execute heuristic folding. To ensure effective standardization, we introduce a novel factorized reward function for each action primitive. This function incorporates garment coverage, keypoint distance, and Io U metrics both before and after each action. This ensures that once unfolded, the garment achieves maximum coverage, maintains consistent orientation, and aligns with its goal shape, while preserving visible keypoints, which are essential for downstream folding tasks. Further, to address the issue of invalid action selections such as targeting areas outside the garment or regions prone to arm collisions we introduce a spatial action mask that filters out actions directed at void regions. Additionally, we present the action optimized module, which enhances unfolding efficiency by focusing on keypoints, particularly the shoulder regions, during the dynamic fling process. To validate the effectiveness of garment standardization in downstream tasks like folding, we combine garment keypoint detection with heuristic folding methods. Additionally, we collect real images of various garment types and create a dedicated dataset for garment folding keypoint detection. Experimental results demonstrate that our method successfully unfolds arbitrarily crumpled garment configurations into standardized states, simplifying the downstream task. The main contributions of this paper are as follows: We propose APS-Net, a self-supervised learning network that intelligently selects between fling and p&p actions. It effectively unfolds garments from arbi- trary crumpled states into standardized configurations, while preserving clear visible keypoints for downstream tasks. We introduce a factorized reward function that combines garment coverage, Io U, and keypoints to guide APS-Net training. We design a spatial action mask to filter out invalid actions and introduce an action optimized module that enhances unfolding efficiency by targeting shoulder keypoints during dynamic fling. We conduct extensive real-world experiments on garments from three categories long sleeve, jumpsuits, and skirts demonstrating the effectiveness of APSNet in both garment standardization and subsequent folding tasks using a dual-UR5 robot. 2. Related work 2.1. Garment unfolding With the advancement of deep learning (Yang et al., 2025a; Bi et al., 2025b; Xin et al., 2024; Li et al., 2024; Yang et al., 2025b; Liu et al., 2022), large language models (LLMs) (Jiang et al., 2025; Tan et al., 2025; Bi et al., 2025a), and reinforcement learning models (Zhou et al., 2025a; Tang et al., 2025; Havrilla et al., 2024), an increasing number of studies have begun to address cloth manipulation within the Learning Efficient Robotic Garment Manipulation with Standardization reinforcement learning framework. Among these, garment unfolding (Wu et al., 2024; Raval et al., 2024) remains a particularly challenging task, requiring intricate manipulation to transform crumpled garments into flat, standardized states. Early approaches relied on hand-crafted heuristics (Seita et al., 2020; Sun et al., 2014) or single-arm manipulation strategies (Qian et al., 2020; Lee et al., 2021), often requiring multiple interactions to achieve successful unfolding. More recent methods, such as those by Zhu et al. (2023) and Chen et al. (2023), employ supervised learning techniques like semantic segmentation and grasping pose estimation to improve unfolding performance. Wu et al. (2023) further introduced topological visual correspondences to handle garments with varying deformations. However, these methods still rely heavily on large labeled datasets and struggle when dealing with heavily wrinkled or deformed garments. Reinforcement learning (RL) provides a promising alternative by enabling robots to optimize unfolding strategies through direct interaction with the environment. Chen et al. (2022) introduced high-velocity dynamic fling actions, but performance plateaus once a surface coverage threshold is reached. Avigal et al. (2022) and He et al. (2024) combined dynamic fling with static pick-and-place actions, but primarily focused on coverage while neglecting standardized alignment. Canberk et al. (2023) explored both single-arm and dual-arm manipulation, though their approach suffers from a significant sim-to-real gap that limits applicability in real-world scenarios. To address these challenges, we propose APS-Net, which ensures efficient garment unfolding with standardized alignment. To reduce the sim-to-real gap, APS-Net uses RGBD input for spatial accuracy, simulates cloth with variable mass and stiffness, and adds procedural wrinkles to enhance realism. 2.2. Garment folding Garment folding(Gu et al., 2024a) is crucial in applications such as hospitals, homes, and warehouses, where efficient fabric manipulation is essential. Early methods relied on traditional image processing to identify garment features(Maitin-Shepard et al., 2010; Doumanoglou et al., 2016; Stria et al., 2014; Ma et al., 2024), but their strong assumptions about garment types and textures limited generalization to real-world scenarios. More recent approaches leverage goal-conditioned policies(Zhou et al., 2024; Chane Sane et al., 2021; Kim et al., 2024) using reinforcement learning, self-supervised learning, and imitation learning. For example, Mo et al. (2022) introduced a space-time attention mechanism for multi-step folding, while Weng et al. (2022) used optical flow techniques to improve performance. However, these methods still depend on predefined goal states, which are often unavailable for novel instances. Recent research has focused on learning-based methods to detect key garment features, such as corners and edges, to reduce task complexity and accommodate diverse garment types. Lips et al. (2024) and Canberk et al. (2023) used synthetic data for training keypoint detection models, achieving effective results, but struggled with the sim-to-real gap in real-world. To address these challenges, we collect a diverse dataset of 10,000 garment states with a coverage rate above 0.8 during ASP-Net training. We then train a keypoint detection model with Deep Labv3(Chen et al., 2017) and fine-tune it using 1,100 real-world images, improving robustness and adaptability in practical scenarios. 3.1. Problem Formulation Standardized garment manipulation involves a sequence of actions that transform a crumpled garment into a standardized, aligned state, followed by folding it into a desired configuration. At each time step t, given an RGB-D image ot RW H 4, an action at of type m M is selected based on the policy πθ, where M is a discrete set of predefined action primitives (See Figure 1). The policy is represented as follows: πθ(ot) at (1) The chosen action at is then applied to the garment, resulting in a new state ot+1, determined by the state transition T : ot+1 T (ot, at) (2) This process continues until the garment reaches a sufficiently smooth state osmooth. After which keypoint detection identifies features for final folding: Ko = Kdetect(osmooth) (3) The garment is then folded based on the detected keypoints using the fold primitive: ofolded = ffold(osmooth, Ko) (4) During this process, we formulate the unfolding process as a self-supervised learning problem, where the policy πθ is learned. The folding process is treated as a supervised learning problem, where Kdetect is learned for keypoint detection. 3.2. Action-Primitive Selector Network (APS-Net) Coarse-grain dynamic dual-arm flings can efficiently unfold garments from crumpled states, but are insufficient for the fine-grained adjustments required to achieve standardized alignment. To address this, we introduce the Action-Primitive Selector Network (APS-Net) (See Figure 2), which intelligently selects the optimal manipulation action fling or p&p for efficient and accurate garment unfolding. Garment unfolding requires precise grasping point Learning Efficient Robotic Garment Manipulation with Standardization Figure 2. Approach overview. APS-Net takes a batch of rotated and scaled RGBD images as input and uses three encoders, each producing a pair of decoders the fling and p&p decoders which are weighted to generate the corresponding Spatial Action Map. A Spatial Action Mask is then applied to filter out invalid actions. The valid primitive batches are concatenated, and the action to be executed is parameterized by the maximum value pixel. Once the garment is sufficiently flattened, a keypoint detection-based method is employed to perform the folding. selection for effective manipulation. For dual-arm fling, directly predicting both grasp points is challenging due to collision avoidance. Instead, we reframe the problem by predicting the midpoint (x, y), angle θ (rotation), and grasp width w, with collision avoidance achieved by adjusting w. For single-arm p&p, collision avoidance is not required. However, to maintain consistency with dual-arm fling and provide a unified output from the same network, we define (x, y) as the pick point, while θ and w represent the rotation and length between the pick and place points. Therefore, we generalize the two cases as solving for x, y, θ, w . Even so, the naive approach of directly predicting x, y, θ, w fails to account for the invariance of the optimal grasp under the garment s physical transformations. To address this, we apply a series of transformations rotation and scaling to the observation ot, generating a set of transformed images Tn to ensure that the grasp points remain consistent across identical garment configurations. APS-Net then processes these transformed garment images in batches to generate spatial action maps representing the desirability of performing each action at each pixel of the garment. For each action primitive m {fling, p&p}, APS-Net computes the maximum desirability across the set of transformed value maps {V(m,1), V(m,2), . . . , V(m,n)} as follows: ˆV m i = max(V m i ) i {1, 2, . . . , n} (5) Here, ˆV m i represents the maximum desirability value found in the i-th transformed value map for action m. The final action primitive is determined by comparing the maximum desirability values from the spatial action maps: ( fling if ˆV fling i > ˆV pick i p&p otherwise (6) Subsequently, the final value V max final is selected based on the chosen action primitive: V max final = ( V(fling,i) if m = fling V(pick,i) if m = p&p (7) The index i is then obtained from the APS-Net output as follows: x, y, i APSNet(Tn) = arg max V max fmax (8) Learning Efficient Robotic Garment Manipulation with Standardization Figure 3. Spatial Action Mask. The series of smaller maps in (a) represents different slices of the spatial action maps for various primitives, each corresponding to a layer at a different scale and rotation. The masks applied to each layer serve to filter out invalid actions those that would cause the robot s end-effector to collide or extend beyond the garment. Once i is determined, the corresponding width w and rotation angle θ are retrieved from the i-th entry of the proposal set Tn: w, θ = Tn[i] (9) Based on the above reasoning, we obtain x, y, θ, w . Note that θ is a discrete variable. As described in Appendix E, θ is quantized into 16 discrete rotation angles to cover the full 360 range. The grasp points at are defined based on the selected action primitive: ( {(x, y + w), (x, y w)} if m = fling {(x, y), (x, y w)} if m = p&p (10) According to the definition (see Figure 1), for single-arm p&p, both the pick and place points need to be predicted. As indicated in Equation (10), when m = p&p, the pick point is (x, y), and the place point is (x, y w). For dual-arm fling, only the two pick points are predicted, while the place points follow a predefined fling trajectory, as described in Equation (11). Specifically, when m = fling, the dual-arm pick points are (x, y + w) and (x, y w). During execution, both arms follow the trajectory through lift, swing, and placement phases, dynamically adjusting acceleration and velocity at each stage to perform the fling motion. fling = [(0, 0, hl) (0, fm, hl) (0, fm bm, hl) (0, fm, hp)] (11) Here, hl is the lift height, fm denotes the forward swing distance, bm is the backward swing offset, and hp is the final placement height. To prevent collisions during dual-arm fling operations, a constraint is imposed to ensure that the left grasp point (L) remains to the left of the right grasp point (R). This is enforced by limiting the rotation angle θ to the range [ 90 , 90 ], ensuring safe and coordinated movements. 3.3. Factorized Reward Function In this work, we propose a novel factorized reward function to train the policy to manipulate the garment into standardized unfolding. APS-Net consists of three encoders, each computing a distinct reward: coverage reward RC, Intersection over Union (Io U) reward RI, and keypoint reward RK. Each encoder is paired with two decoder heads one for each action primitive, fling and p&p. These rewards are computed separately for each action primitive and then combined into a weighted sum to guide network training. To balance the contributions of these reward functions, a weighted sum is applied for each action primitive: RCIK = α RC + β RI + (1 α β) RK (12) where α, β (0, 1) are hyperparameters. By tuning α and β, this reward mechanism enables APS-Net to optimize fabric coverage, alignment, and keypoint positioning, achieving a standardized garment configuration critical for subsequent folding tasks. See Appendix A for more details. 3.4. Spatial Action Mask Directly sampling actions on the spatial policy maps can lead to invalid actions, such as empty actions in non-garment areas or double-arm collisions. To address this, we introduce the Spatial Action Mask (SAM), which filters out infeasible actions, ensuring that the robot only considers feasible Learning Efficient Robotic Garment Manipulation with Standardization Figure 4. The folding rules for long sleeves, pants, and skirts. Figure 5. Action Optimized Module. Optimizes the alignment of garment shoulder keypoints for efficient flattening, based on the predicted shoulder points from APSNet. actions within the valid workspace (See Figure 3). Specifically, we define both workspace and garment masks for each action primitive, and the valid action space is determined by their intersection. See Appendix B for more details. 3.5. Keypoint-Based Strategy for Garment Folding Heuristics are highly interpretable and easy to define, making them effective for garment folding. In this approach, a Deep Labv3 detector(Chen et al., 2017) is trained for each garment class, and keypoints from the detected garments are depth-projected into 3D space, then transformed into the workspace frame of reference. By representing garment as a set keypoints, we sidestep its infinite Do F by using a few meaningful keypoints as the representation, which makes it simpler to define heuristics over them. For long-sleeve garments, we define six key points: two at the sleeves, two at the shoulders, and two at the waist. First, fold the sleeves along the folding lines (Figure 4(a)), then fold the garment in half along the central line. For pants and skirts (Figure 4(b) and 4(c)), we define four key points: two at the waist and two at the bottom. Begin by folding along the waist and bottom lines, then fold 3/4 of the waist down toward the bottom key point to complete the fold. 3.6. Action Optimized Module When performing a fling action to flatten a garment, keypoints on the shoulders are typically prioritized for efficient flattening. The Action Optimized Module (AOM) refines the APSNet-predicted points P1 and P2 by aligning them with the garment s shoulder keypoints K1 and K2, if they are sufficiently close, ensuring quick flattening(See Figure 5). The module calculates the Euclidean distance between each predicted point and the corresponding keypoint: d(Pi, Kj) = q (Pi,x Kj,x)2 + (Pi,y Kj,y)2 (13) where i, j {1, 2}. Two possible matching combinations are evaluated, with p1 paired with k1 and p2 paired with k2, which results in the following: D1 = d(P1, K1) + d(P2, K2) (14) and pairing p1 with k2 and p2 with k1, yielding: D2 = d(P1, K2) + d(P2, K1) (15) The module selects the pair with the minimum total distance, min(D1, D2). If both distances in the selected pair satisfy d(Pi, Kj) < δ, where δ = 5, the predicted points P1 and P2 are replaced by K1 and K2. Otherwise, they are retained. 4. EXPERIMENTS 4.1. Tasks and Metrics Intersection over Union (Io U): Measures the quality of Garment Unfold by comparing the manipulated Garment s mask with the target mask. Coverage (Cov): Compares initial and achieved Garment coverage to the maximum possible area. Success Rate (SR): A task is considered successful if the robot first smooths the crumpled garment and then correctly executes the folding operation according to the specified fold primitive. In simulation: The cloth is modeled as a particle-based grid, where ground-truth particle positions are available. We compute the mean distance between particles in the achieved and target cloth states. A trial is deemed successful if the average particle distance error is less than 0.03. In the real world: We use the Intersection over Union (Io U) between the predicted cloth mask and that of a human demonstrator to evaluate performance. If the Io U exceeds 0.8, the folding is considered successful. Keypoint distance (KD): Represents the distance between key feature points of the garment. These metrics are applied to three representative garment manipulation tasks: unfolding and folding long sleeves, skirts, and pants. In all tables, indicates that lower is better while indicates that larger values are preferred. 4.2. Baselines We compare our method with several baselines, including quasi-static, dynamic, and state-of-the-art (SOTA) ap- Learning Efficient Robotic Garment Manipulation with Standardization Figure 6. Qualitative comparison of simulation unfolding. Our method outperforms the four other methods. The red arrow denotes p&p action, while the green line indicates a fling action. The gray contour line represents the standardized garment template. Table 1. Results of our method and four baselines. TYPE COV(%) IOU(%) KD RCIK P&P 61.3 51.6 2.759 0.261 S-FLING 80.5 63.3 1.892 0.135 FLINGBOT 80.1 60.9 2.084 0.086 CLOTH FUNNLES 87.2 74.1 1.972 0.058 OURS 91.1 79.3 1.832 0.042 proaches. Quasi-static Methods include p&p, where the garment is sequentially lifted and placed, similar to (Lee et al., 2021). Dynamic Methods consist of Fling Bot(Ha & Song, 2022), which uses dual-arm dynamic flinging with coverage as the reward function, and S-fling, a variant of our method using a single fling action. SOTA Methods include Cloth funnels(Canberk et al., 2023), which combines p&p and fling actions, utilizing factorized distances between cloth particles for unfolding. 4.3. Results and Analysis Due to the complex shape and high manipulation difficulty of long sleeves garments, we selected them as the target for our experiments. Each policy was evaluated through 50 trials with random initial configurations of the garments. For unfolding, Table 1 shows the quantitative results, while Figure 7. Garment Unfolding Coverage and Io U vs. Steps. Figure 6 and Figure 7 present representative qualitative results. A comparison between S-fling, Flingbot, and p&p reveals that the former two methods significantly improve coverage due to the dynamic nature of the fling actions. However, while coarse-grain high-velocity actions are effective for unfolding, they are not versatile enough for standardization. Cloth funnels outperform S-fling in coverage and Io U by facilitating finer adjustments with quasi-static actions. However, its reward function, based on the 3D distance between cloth particles, can lead to local minima during policy learning, limiting its ability to achieve optimal unfolding. In contrast, our method introduces a novel 2D plane-based factorized reward function, which achieves higher coverage and Io U with fewer manipulation steps, ensures garment standardization, and demonstrates the synergy between coarse-grain high-velocity flings and fine-grain p&p actions for optimal performance. Learning Efficient Robotic Garment Manipulation with Standardization Figure 8. Unfolding results for different garment types. Table 2. Performance metrics for Pants and Skirts. TYPE COV(%) IOU(%) KD PANTS 91.8 80.4 1.819 SKIRTS 86.7 79.6 1.825 We trained models on two additional garment categories pants and skirts using the same configuration(See Figure 8). As shown in Table 2, our method achieves around 80% Io U across all categories, demonstrating its generalizability. Figure 9. Simulation Folding Qualitative Comparison. We compare the unfolding results of our method with those of other existing methods using our folding heuristic. For folding, we compare our method with existing approaches. Our method achieves superior folding performance(See Figure 9). Unlike methods focused solely on maximizing coverage, standardized unfolding aligns the garment s shape and keypoints, which simplifies the folding process and making it well-suited for garment packing in factory settings. In our experiments, we set the hyperparameters to α = 0.2 and β = 0.4. To assess the sensitivity of our method to these parameters, we conducted a series of experiments with different values of α and β (See Table 3). Table 3. Effect of different hyperparameters α and β. PARAMETER COV(%) IOU(%) KD α = 0.2, β = 0.2 89.9 76.5 1.844 α = 0.2, β = 0.4 91.1 79.3 1.832 α = 0.4, β = 0.2 90.7 78.6 1.841 α = 0.4, β = 0.4 89.4 77.5 1.863 α = 0.6, β = 0.2 88.7 77.9 1.869 Table 4. Ablation Studies. METHOD COV(%) IOU(%) KD W/O KPR 88.2 76.4 1.851 W/O AOM 89.5 77.6 1.860 W/O IR 90.3 72.2 1.994 OURS 91.1 79.3 1.832 4.4. Ablation Studies To evaluate the importance of different components in our method, we conduct ablation experiments by comparing our method with the following, with 50 trials per ablation using random configurations of long sleeves garments: W/O AOM: Excludes the Action Optimized Module, resulting in no optimization for the fling action. W/O IR: Excludes the Io U Reward, which helps the garment achieve better coverage, potentially reducing accuracy in spatial matching. W/O KPR: Removes the Key Point Reward, which directly aids in keypoint detection, potentially impairing performance in precise keypoint localization. Table 4 shows quantitative comparisons with ablations. Clearly, each component improves our method s capability. Table 5. Results of real world experiments on Garment Unfolding. METHOD IOU (%) COV (%) P&P 43.7 52.3 FLINGBOT 51.3 72.8 CLOTH FUNNELS 65.2 82.2 OURS 70.5 86.3 4.5. Real World Results In our real-world experiments, we performed unfolding and folding tasks on long sleeve garments, conducting 15 trials for each group, with models trained in a simulated environment. For the unfolding task, quantitative results are presented in Table 5. Our method achieved the best performance, reaching 70.5% in Io U and 86.3% in Cov, while preserving the standardization of garment , thus validating the conclusions drawn from the simulation. Learning Efficient Robotic Garment Manipulation with Standardization Figure 10. Real-world Folding. Figure 11. Real-world Folding in pants and skirts. Table 6. Success rates across different methods. METHOD SUCCESS RATE P&P 1/15 FLINGBOT 6/15 CLOTH FUNNELS 10/15 OURS 12/15 We applied our keypoint-based folding model to various unfolding methods, as shown in Table 6 (quantitative results) and Figure 10 (qualitative results). Our model achieved a success rate of 12 out of 15, demonstrating superior performance compared to existing methods. Furthermore, the results demonstrate that standardization reduces the complexity of downstream tasks. Additionally, we conducted experiments on pants and skirts Table 7. Results of real world experiments on other garment types. METHOD SLEEVE PANTS SKIRT FOLD 12/15 14/15 11/15 (See Figure 11 and Table 7). The results demonstrate that our method generalizes to different types of garments. 5. Conclusion In this paper, we presented APS-Net, a novel approach for garment manipulation that integrates both unfolding and standardization into a unified framework. By leveraging a dual-arm, multi-primitive policy, APS-Net efficiently unfolds crumpled garments and ensures precise standardization, which simplifies downstream tasks like folding. We introduced a factorized reward function, spatial action mask, and an action optimized module to enhance performance. Experimental results show that APS-Net outperforms existing methods in both simulation and real-world folding tasks, demonstrating its potential for practical garment manipulation. Our approach provides a solid foundation for further advancements in robot-assisted garment handling for a range of applications. Learning Efficient Robotic Garment Manipulation with Standardization Impact Statement This paper presents a method for garment manipulation. It has potential applications in household and industrial automation, reducing manual labor. We have not identified any particular ethical issues that need to be emphasized. Acknowledgements This work was supported in part by the National Natural Science Foundation of China (No. 62088101), in part by the Science and Technology Commission of Shanghai Municipality (No. 2021SHZDZX0100, 22ZR1467100), and the Fundamental Research Funds for the Central Universities (No. 22120240291). Arents, J. and Greitans, M. Smart industrial robot control trends, challenges and opportunities within manufacturing. Applied Sciences, 12(2):937, 2022. Avigal, Y., Berscheid, L., Asfour, T., Kr oger, T., and Goldberg, K. Speedfolding: Learning efficient bimanual folding of garments. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 1 8. IEEE, 2022. Bertiche, H., Madadi, M., and Escalera, S. Cloth3d: clothed 3d humans. In European Conference on Computer Vision, pp. 344 359. Springer, 2020. Bi, J., Wang, Y., Chen, H., Xiao, X., Hecker, A., Tresp, V., and Ma, Y. Llava steering: Visual instruction tuning with 500x fewer parameters through modality linear representation-steering, 2025a. URL https://arxiv. org/abs/2412.12359. Bi, J., Wang, Y., Yan, D., Xiao, X., Hecker, A., Tresp, V., and Ma, Y. Prism: Self-pruning intrinsic selection method for training-free multimodal data selection, 2025b. URL https://arxiv.org/abs/2502.12119. Brockman, G., Cheung, V., Pettersson, L., Schneider, J., Schulman, J., Tang, J., and Zaremba, W. Openai gym, 2016. Canberk, A., Chi, C., Ha, H., Burchfiel, B., Cousineau, E., Feng, S., and Song, S. Cloth funnels: Canonicalizedalignment for multi-purpose garment manipulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pp. 5872 5879. IEEE, 2023. Chane-Sane, E., Schmid, C., and Laptev, I. Goalconditioned reinforcement learning with imagined subgoals. In International conference on machine learning, pp. 1430 1440. PMLR, 2021. Chen, L.-C., Papandreou, G., Schroff, F., and Adam, H. Rethinking atrous convolution for semantic image segmentation, 2017. URL https://arxiv.org/abs/ 1706.05587. Chen, L. Y., Huang, H., Novoseller, E., Seita, D., Ichnowski, J., Laskey, M., Cheng, R., Kollar, T., and Goldberg, K. Efficiently learning single-arm fling motions to smooth garments. In The International Symposium of Robotics Research, pp. 36 51. Springer, 2022. Chen, V. K., Chin, L., Choi, J., Zhang, A., and Rus, D. Real-time grocery packing by integrating vision, tactile sensing, and soft fingers. In 2024 IEEE 7th International Conference on Soft Robotics (Robo Soft), pp. 392 399. IEEE, 2024. Chen, W., Lee, D., Chappell, D., and Rojas, N. Learning to grasp clothing structural regions for garment manipulation tasks. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 4889 4895. IEEE, 2023. Doumanoglou, A., Stria, J., Peleka, G., Mariolis, I., Petrik, V., Kargakos, A., Wagner, L., Hlav aˇc, V., Kim, T.-K., and Malassiotis, S. Folding clothes autonomously: A complete pipeline. IEEE Transactions on Robotics, 32 (6):1461 1478, 2016. Gu, N., He, R., and and, L. Y. Defnet: deconstructed strategy for multi-step fabric folding tasks. The Journal of The Textile Institute, 0(0):1 9, 2024a. doi: 10.1080/ 00405000.2024.2400633. URL https://doi.org/ 10.1080/00405000.2024.2400633. Gu, N., He, R., and Yu, L. Learning to unfold garment effectively into oriented direction. IEEE Robotics and Automation Letters, 9(2):1051 1058, 2024b. doi: 10. 1109/LRA.2023.3341763. Ha, H. and Song, S. Flingbot: The unreasonable effectiveness of dynamic manipulation for cloth unfolding. In Conference on Robot Learning, pp. 24 33. PMLR, 2022. Havrilla, A., Du, Y., Raparthy, S. C., Nalmpantis, C., Dwivedi-Yu, J., Zhuravinskyi, M., Hambro, E., Sukhbaatar, S., and Raileanu, R. Teaching large language models to reason with reinforcement learning. ar Xiv preprint ar Xiv:2403.04642, 2024. He, C., Meng, L., Sun, Z., Wang, J., and Meng, M. Q.-H. Fabricfolding: learning efficient fabric folding without expert demonstrations. Robotica, 42(4):1281 1296, 2024. Hoque, R., Shivakumar, K., Aeron, S., Deza, G., Ganapathi, A., Wong, A., Lee, J., Zeng, A., Vanhoucke, V., and Goldberg, K. Learning to fold real garments with one arm: A case study in cloud-based robotics research. In Learning Efficient Robotic Garment Manipulation with Standardization 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 251 257, 2022. doi: 10.1109/IROS47612.2022.9981253. Huang, Z., Lin, X., and Held, D. Mesh-based dynamics with occlusion reasoning for cloth manipulation. ar Xiv preprint ar Xiv:2206.02881, 2022. Jiang, Y., Gao, X., Peng, T., Tan, Y., Zhu, X., Zheng, B., and Yue, X. Hiddendetect: Detecting jailbreak attacks against large vision-language models via monitoring hidden states, 2025. URL https://arxiv.org/abs/ 2502.14744. Jing, P., Cui, K., Guan, W., Nie, L., and Su, Y. Categoryaware multimodal attention network for fashion compatibility modeling. IEEE Transactions on Multimedia, 25: 9120 9131, 2023. Jing, P., Cui, K., Zhang, J., Li, Y., and Su, Y. Multimodal high-order relationship inference network for fashion compatibility modeling in internet of multimedia things. IEEE Internet of Things Journal, 11(1):353 365, 2024. Kim, H., Ohmura, Y., and Kuniyoshi, Y. Goal-conditioned dual-action imitation learning for dexterous dual-arm robot manipulation. IEEE Transactions on Robotics, 2024. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A. C., Lo, W.-Y., et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 4015 4026, 2023. Lee, R., Ward, D., Dasagi, V., Cosgun, A., Leitner, J., and Corke, P. Learning arbitrary-goal fabric folding with one hour of real robot experience. In Conference on Robot Learning, pp. 2317 2327. PMLR, 2021. Li, S., Li, B., Sun, B., and Weng, Y. Towards visual-prompt temporal answer grounding in instructional video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):8836 8853, 2024. doi: 10.1109/TPAMI. 2024.3411045. Li, Y., Wu, J., Tedrake, R., Tenenbaum, J. B., and Torralba, A. Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids. ar Xiv preprint ar Xiv:1810.01566, 2018. Lin, X., Wang, Y., Olkin, J., and Held, D. Softgym: Benchmarking deep reinforcement learning for deformable object manipulation. In Conference on Robot Learning, pp. 432 448. PMLR, 2021. Lin, X., Wang, Y., Huang, Z., and Held, D. Learning visible connectivity dynamics for cloth smoothing. In Conference on Robot Learning, pp. 256 266. PMLR, 2022. Lips, T., De Gusseme, V.-L., et al. Learning keypoints for robotic cloth manipulation using synthetic data. IEEE Robotics and Automation Letters, 2024. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. ar Xiv preprint ar Xiv:2303.05499, 2023. Liu, W., Wang, H., Wang, J., Li, R., Yue, C., and Zhang, Y. Fr: Folded rationalization with a unified encoder, 2022. URL https://arxiv.org/abs/2209.08285. Ma, X., Patidar, S., Haughton, I., and James, S. Hierarchical diffusion policy for kinematics-aware multi-task robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 18081 18090, 2024. Maitin-Shepard, J., Cusumano-Towner, M., Lei, J., and Abbeel, P. Cloth grasp point detection based on multipleview geometric cues with application to robotic towel folding. In 2010 IEEE International Conference on Robotics and Automation, pp. 2308 2315. IEEE, 2010. Mo, K., Xia, C., Wang, X., Deng, Y., Gao, X., and Liang, B. Foldsformer: Learning sequential multi-step cloth manipulation with space-time attention. IEEE Robotics and Automation Letters, 8(2):760 767, 2022. Moglia, A., Georgiou, K., Cerveri, P., Mainardi, L., Satava, R. M., and Cuschieri, A. Large language models in healthcare: from a systematic review on medical examinations to a comparative analysis on fundamentals of robotic surgery online test. Artificial Intelligence Review, 57(9):231, 2024. Qian, J., Weng, T., Zhang, L., Okorn, B., and Held, D. Cloth region segmentation for robust grasp selection. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 9553 9560. IEEE, 2020. Raval, V., Zhao, E., Zhang, H., Nikolaidis, S., and Seita, D. Gpt-fabric: Folding and smoothing fabric by leveraging pre-trained foundation models. ar Xiv preprint ar Xiv:2406.09640, 2024. Salhotra, G., Liu, I., Arthur, C., and Sukhatme, G. Learning robot manipulation from cross-morphology demonstration. ar Xiv preprint ar Xiv:2304.03833, 2023. Seita, D., Ganapathi, A., Hoque, R., Hwang, M., Cen, E., Tanwani, A. K., Balakrishna, A., Thananjeyan, B., Ichnowski, J., Jamali, N., et al. Deep imitation learning of sequential fabric smoothing from an algorithmic supervisor. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 9651 9658. IEEE, 2020. Learning Efficient Robotic Garment Manipulation with Standardization Stria, J., Prˇsa, D., Hlav aˇc, V., Wagner, L., Petrik, V., Krsek, P., and Smutn y, V. Garment perception and its folding using a dual-arm robot. In 2014 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 61 67. IEEE, 2014. Sun, L., Aragon-Camarasa, G., Cockshott, P., Rogers, S., and Siebert, J. P. A heuristic-based approach for flattening wrinkled clothes. In Towards Autonomous Robotic Systems: 14th Annual Conference, TAROS 2013, Oxford, UK, August 28 30, 2013, Revised Selected Papers 14, pp. 148 160. Springer, 2014. Tan, Y., Jiang, Y., Li, Y., Liu, J., Bu, X., Su, W., Yue, X., Zhu, X., and Zheng, B. Equilibrate rlhf: Towards balancing helpfulness-safety trade-off in large language models, 2025. URL https://arxiv.org/abs/2502. 11555. Tang, C., Abbatematteo, B., Hu, J., Chandra, R., Mart ın Mart ın, R., and Stone, P. Deep reinforcement learning for robotics: A survey of real-world successes. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp. 28694 28698, 2025. Wang, Y., Zhong, J., and Kumar, R. A systematic review of machine learning applications in infectious disease prediction, diagnosis, and outbreak forecasting. 2025. Weng, T., Bajracharya, S. M., Wang, Y., Agrawal, K., and Held, D. Fabricflownet: Bimanual cloth manipulation with a flow-based policy. In Conference on Robot Learning, pp. 192 202. PMLR, 2022. Wu, R., Ning, C., and Dong, H. Learning foresightful dense visual affordance for deformable object manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 10947 10956, 2023. Wu, R., Lu, H., Wang, Y., Wang, Y., and Dong, H. Unigarmentmanip: A unified framework for category-level garment manipulation via dense visual correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 16340 16350, 2024. Xiao, A., Gupta, A., Deng, Y., Li, K., and Hsu, D. Robi butler: Multimodal remote interaction with household robotic assistants. In 2nd Workshop on Mobile Manipulation and Embodied Intelligence at ICRA 2024. Xin, Y., Du, J., Wang, Q., Yan, K., and Ding, S. Mmap: Multi-modal alignment prompt for cross-domain multitask learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp. 16076 16084, 2024. Xue, H., Li, Y., Xu, W., Li, H., Zheng, D., and Lu, C. Unifolding: Towards sample-efficient, scalable, and generalizable robotic garment folding. ar Xiv preprint ar Xiv:2311.01267, 2023. Yang, J., Deng, C., Wu, J., Antonova, R., Guibas, L., and Bohg, J. Equivact: Sim (3)-equivariant visuomotor policies beyond rigid object manipulation. In 2024 IEEE international conference on robotics and automation (ICRA), pp. 9249 9255. IEEE, 2024a. Yang, L., Li, Y., and Chen, L. Clothppo: A proximal policy optimization enhancing framework for robotic cloth manipulation with observation-aligned action spaces. ar Xiv preprint ar Xiv:2405.04549, 2024b. Yang, S., Luo, S., and Han, S. C. Multimodal commonsense knowledge distillation for visual question answering (student abstract). In Proceedings of the AAAI conference on artificial intelligence, volume 39, pp. 29545 29547, 2025a. Yang, S., Luo, S., Han, S. C., and Hovy, E. Magic-vqa: Multimodal and grounded inference with commonsense knowledge for visual question answering. ar Xiv preprint ar Xiv:2503.18491, 2025b. Zhang, F. and Demiris, Y. Learning garment manipulation policies toward robot-assisted dressing. Science robotics, 7(65):eabm6010, 2022. Zhou, C., Jiang, R., Luan, F., Meng, S., Wang, Z., Dong, Y., Zhou, Y., and He, B. Dual-arm robotic fabric manipulation with quasi-static and dynamic primitives for rapid garment flattening. IEEE/ASME Transactions on Mechatronics, pp. 1 11, 2025a. doi: 10.1109/TMECH. 2025.3556283. Zhou, C., Xu, H., Hu, J., Luan, F., Wang, Z., Dong, Y., Zhou, Y., and He, B. Ssfold: Learning to fold arbitrary crumpled cloth using graph dynamics from human demonstration. IEEE Transactions on Automation Science and Engineering, 22:14448 14460, 2025b. doi: 10.1109/TASE.2025.3560680. Zhou, P., Qi, J., Duan, A., Huo, S., Wu, Z., and Navarro Alarcon, D. Imitating tool-based garment folding from a single visual observation using hand-object graph dynamics. IEEE Transactions on Industrial Informatics, 20(4): 6245 6256, 2024. doi: 10.1109/TII.2023.3342895. Zhu, X., Wang, X., Freer, J., Chang, H. J., and Gao, Y. Clothes grasping and unfolding based on rgb-d semantic segmentation. In 2023 IEEE International conference on robotics and automation (ICRA), pp. 9471 9477. IEEE, 2023. Learning Efficient Robotic Garment Manipulation with Standardization A. Factorized Reward Function Details The Coverage Reward RC measures the proportion of newly covered garment area relative to the total garment area, promoting actions that maximize garment coverage: RC = C Cmax (16) where C is a function that computes the current area of the cloth based on the image pixels, and Cmax is the maximum area of the garment. The Io U Reward RI evaluates the overlap between the current garment configuration mc and the target configuration mgt, ensuring proper alignment: RI = |mc mgt| |mc mgt| (17) The Keypoint Reward RK is designed to evaluate the accuracy of key garment feature positioning by minimizing the distance between the predicted and target keypoints. i=1 pi p i (18) where pi represents the current position of keypoint i, p i is its target position, and N is the total number of keypoints. B. Spatial Action Mask Details Directly sampling actions directly on the existing spatial policy maps will lead to many invalid actions, such as empty actions in parts other than garments. The shape of the garment mask varies due to differences in rotation and scaling. To address this, we introduce the Spatial Action Mask (SAM) to filter out invalid actions, ensuring that the robot only considers feasible actions within its workspace. The robot s arm reach is constrained by its physical limits, which are represented by masks in the simulation. The workspace is defined by a table width w = 1.53 m, with the left and right arm bases at positions [0.765, 0, 0] and [ 0.765, 0, 0], respectively. The maximum reachable distance of a single arm is denoted as rd. The pixel radius rpix is calculated by scaling rd relative to w, as: rpix = D rd/w (19) where D is the rendered image dimension. Circular masks for both arms are created, with the left arm centered at the upper half and the right arm at the lower half of the image. These masks are drawn in white, indicating reachable areas (see Figure 12 (a), (b)). Workspace Mask. Both the p&p and fling actions rely on workspace masks to filter the valid action space for the robot. For the p&p action, only one arm is used at a time. The valid action space is determined by the union of the workspace masks for the left and right arms. To account for each arm s reach, the workspace masks for both arms are each shifted down by w pixels, as described in formula (9) (see Figure 12(c), (d)). The final p&p workspace mask m(p&p) w is obtained by taking the union of the shifted masks (see Figure 12(e)). For the fling action, both arms are used simultaneously. The valid action space is determined based on the position of the grabbing point p1. If p1 corresponds to the left arm, the left arm s workspace mask is shifted upwards by w pixels, and the right arm s workspace is shifted downwards by w pixels. The valid action space is the intersection of these two shifted masks (see Figure 12(j)). If p1 corresponds to the right arm, the left arm workspace mask is shifted downwards by w pixels, and the right arm workspace mask is shifted upwards by w pixels. The valid action space is the intersection of these two shifted masks (see Figure 12(k)). The final fling workspace arm mask mf w is obtained by taking the union of these two intersection masks (see Figure 12(l)). The workspace mask for the action Mw mask can thus be represented as: ( m(p&p) w if m = p&p mf w if m = fling (20) Garment mask. The valid garment mask for different action primitives is computed as follows. In the p&p action, since the grabbing points are directly on the garment, the valid garment mask is simply the original garment mask mg. For the fling action, the grabbing points for both arms are determined using formula (9). To ensure that these points lie within the garment mask, the original garment mask is shifted downward by w pixels for the left arm and upward by w pixels for the right arm. These shifted masks are shown in green and blue (see Figure 13(c)). The final fling garment mask mf g is obtained by taking the intersection of the original garment mask with both the upward and downward shifted masks (see Figure 13(d)). Therefore, the garment mask for action Mg mask is: ( mg m = p&p mf g if m = fling (21) The valid action space is determined by the intersection of the garment mask Mg mask and the workspace mask Mw mask: Mt = Mg mask Mw mask (22) This combined mask Mt filters the action probabilities output by APS-Net. If an action at falls within the valid region (Mt(at) = 1), it is retained; otherwise, its probability is set Learning Efficient Robotic Garment Manipulation with Standardization Figure 12. Dual arm workspace for different action primitives. Figure 13. Garment for different action primitives. to , excluding it from consideration. The final action probability π θ(at|ot) is: π θ(at|ot) = ( πθ(at|ot) if Mt(at) = 1, if Mt(at) = 0. (23) C. Experiment Setup. C.1. Simulation Setup We built upon the Open AI Gym API(Brockman et al., 2016) and Py Fle X(Li et al., 2018) bindings to Nvidia Fle X, integrated with Soft Gym(Lin et al., 2021), to develop a reinforcement learning environment called Cloth Action Gym, which supports loading arbitrary cloth meshes, such as Tshirts, pants, and skirts, through a Python API. The robot gripper is modeled as a spherical picker that can move freely in 3D space and can be activated to attach to the nearest particle. Observations are rendered in Blender 3D, with the cloth s color uniformly sampled across specified HSV ranges. C.2. Real World Setup Our experimental setup comprises two 6-Do F UR5 robot arms, each equipped with a Robotiq 2F-85 gripper, positioned 1.45 meters apart and facing each other. Two topdown RGB-D cameras an Azure Kinect and a Realsense D455 capture the workspace. The Grounded-SAM model (Liu et al., 2023; Kirillov et al., 2023) is used to segment the garment from the RGB image with the prompt garment. To minimize the domain gap between simulation and realworld images, we apply an Affine Transform to crop and correct the camera s perspective, focusing on the central region of the garment. The resulting mask is then multiplied by the RGBD image to generate the input for the model. (See Figure 14) This approach allows the simulation-trained model to be effectively adapted for real-world use without requiring additional training. D. Data Collection In the simulation environment, we collected two datasets, one for training garment unfolding and another for training garment keypoint detection. Each task is characterized by specific parameters such as the cloth mesh, mass, stiffness, and initial configuration. The cloth meshes were sampled from a subset of shirts in the test split of the CLOTH3D dataset(Bertiche et al., 2020), resized to ensure they fall within the robot s operational range. The cloth mass was selected from a range of [0.2 kg, 2.0 kg], and internal stiffness was set between [0.85 kg/s2, 0.95 kg/s2]. To initial- Learning Efficient Robotic Garment Manipulation with Standardization Figure 14. Real-world Setup. ize the cloth states, we randomly sample from a subset of CLOTH3D, which provides a curated collection of garment meshes. For the garment unfolding dataset, the initial configuration was created by randomly rotating the garment, selecting a random point on the garment, and dropping it from a height between [0.6 m, 1.6 m]. The garment was then translated by a random distance in the range of [0 m, 0.25 m], resulting in a highly crumpled configuration. Each garment category includes 2000 training tasks and 50 testing tasks using unseen garment meshes. For the garment keypoint detection dataset, we filtered images from the replay buffers generated during the unfold task, selecting images where the coverage exceeded 0.8. This process resulted in a dataset of 10000 images for training and 500 for testing. Real-world data collection involved 1100 images, which were labeled using Label Me. E. Training Details For the unfolding network, the initial observation ot of size (H, W) = (128, 128) undergoes 16 rotations (covering 360 ) and 5 scales {0.75, 1.0, 1.5, 2.0, 2.5}, yielding m = 80 stacked transformed observations. Exploration follows a decaying ϵ-greedy strategy (initial ϵ = 1) with half-lives of 5000 steps for selecting action primitives (fling vs. p&p) . Optimization is conducted using the Adam optimizer with a learning rate of 1.0 10 3 over 100,000 steps. The value network is supervised using delta-reward values, defined as the difference in weighted reward ( RCIK) before and after each action: RCIK = RCIK,after RCIK,before (24) The value network minimizes the Mean Squared Error (MSE) loss between the ground truth delta-reward Rgt CIK and the predicted delta-reward RCIK: i=1 ( Rgt CIK RCIK)2 (25) For the key-point detection model, training is conducted for 200 epochs with a batch size of 32, a learning rate of 1.0 10 4, and includes rotation and scaling augmentations. The training loss L(y, p) is defined as the mean point-wise cross-entropy loss for each keypoint n in N: L(y, p) = 1 yn log(pn) + (1 yn) log(1 pn) (26) where yn is the true label and pn is the predicted probability for each keypoint n. All experiments are conducted on an NVIDIA RTX 4090 GPU with an Intel i9-13900K CPU (5.80 GHz), supported by 64 GB RAM on Ubuntu 18.04 LTS. F. Experimental Results Figure 15 shows how our method transfers the garment from an arbitrary configuration to a standardized one. Figure 16 presents the qualitative results of garment unfolding in the Learning Efficient Robotic Garment Manipulation with Standardization Figure 15. Our method transfers the garment from an arbitrary configuration to a standardized configuration. The first row shows the spatial action map output by the APS-Net network at each step. The second row maps the highest pixels of the spatial action map to the transformed image at each step, producing the corresponding actions. The third row applies rotation and scaling at each step to transform these actions back to the original input image. real world, comparing our method with other approaches. In Figure 17, we compare the unfolding results of our method with those of Cloth Funnels using our folding heuristic. Finally, Figure 18 compares optimized and unoptimized shoulder keypoint alignment for garment flattening. Figure 16. Qualitative results of garment unfolding in real world. Our method outperforms three other methods on long sleeve garments with four different colors and textures. Figure 17. We compare the unfolding results of our method with those of Cloth Funnels using our folding heuristic. Figure 18. Comparison of Optimized vs. Unoptimized Shoulder Key Point Alignment for Garment Flattening.