Due to differences in learning paradigms, the types of data collected vary across studies. Visual information, such as RGB images, depth images, and grayscale images, is collected in all selected papers. In addition to visual observations, other data modalities are also used to provide richer supervision. These include keypoint annotations (Seita et al., 2019; Hoque et al., 2022b) for keypoint perception, action-related data such as pick and place points (Tsurumine et al., 2019), pick-and-pull directions (Seita et al., 2020; Hoque et al., 2022a), and manipulation trajectories (Chen et al., 2022) for supervising control policy learning, as well as manipulation stage or phase annotations (Mo et al., 2022; Wang et al., 2022) to support temporal decomposition of manipulation processes. Another category of collected data is cloth state representations, such as cloth particle poses (Weng et al., 2022) and cloth mesh representations (Tanaka et al., 2021; Ma et al., 2022), which allow explicit representations of cloth dynamics and deformation. A single study may collect multiple types of data depending on its learning formulation.

Collecting datasets or training in simulation is a widely used strategy for learning-based robotic cloth manipulation (65.9%, 27 out of 41). The most commonly used simulators in this field include gym-based environments, such as SoftGym (Lin et al., 2021), which accounts for more than half of the simulation studies, and FEM-based simulators integrated with Gym (Seita et al., 2020), as well as Blender, MuJoCo, and others. SoftGym (Lin et al., 2021), built on the PyFleX (Li et al., 2019) bindings to NVIDIA FleX, can load various garment meshes including T-shirts, trousers, and dresses. However, its current version does not support loading robot models due to NVIDIA’s permission constraints, limiting its use for training real robots that rely on Cartesian control. FEM-based fabric simulators interfaced with OpenAI Gym provide another option, though the authors acknowledge that these simulators exhibit lower physical fidelity compared to other engines. Hietala et al. (2022) instead used MuJoCo, which supports loading robot URDF models. All simulation datasets mentioned above were collected entirely within simulated environments. Xue et al. (2023) employed RFUniverse (Fu H. et al., 2023), which enables the acquisition of cloth-environment interaction data through a Virtual Reality (VR) setup, thereby enhancing interactive capabilities from the real world to the simulated environment.

While simulation facilitates the development of DL-based methods, the gap between simulation and the real world remains a challenge in robotic cloth manipulation. This gap mainly comes from inaccurate modeling of cloth properties, limited visual diversity, dynamics mismatch, and unmodeled interactions, etc. Therefore, addressing the Sim2Real gap is an important consideration (Collins et al., 2019; Matas et al., 2018). The strategies include Domain Randomization (DR), Data Augmentation (DA), depth or point-cloud observations, fine-tuning with real-world data, texture replacement, and training under real settings, and others. Table 2 summarizes the primary Sim2Real gaps encountered in cloth manipulation and the corresponding strategies used to mitigate them.

Collecting datasets or training in the real world presents challenges but offers significant benefits. It avoids the Sim2Real problem and generally results in better generalization compared to simulation-based training. However, the considerable workload of human annotation and the wear and tear on robots are notable factors. To reduce the labeling workload, Seita et al. (2019) proposed a color-based keypoint annotation strategy that enables automatic extraction of cloth corners from color-marked visual observations for depth images. Later, Fu T. et al. (2023) adopted a similar color-labeling strategy for dataset collection and segmentation training, using different paint color for cloth corners and edges. However, the training dataset collected using this approach includes only depth images, which consist of a single depth channel and contain no color cues. Seita et al. (2020) reported that the color contrast between the fabric and the workspace in RGB images can facilitate better performance. Furthermore, depth sensing requires dedicated hardware. Therefore, Thananjeyan et al. (2022) addressed this by introducing a UV-based labeling technique for deformable objects in RGB images, referred to as Labels from UltraViolet (LUV). Transparent UV fluorescent paint is invisible under standard light but detectable under UV light. Similarly, Gu et al. (2024) adopted this method during their finetuning process.

The color-labeling methods mentioned above can efficiently provide keypoints for cloth manipulation. However, they may be inadequate in certain scenarios, particularly when RL or imitation learning is employed. Consequently, other data collection methods have been explored. For example, Hoque et al. (2022b) developed open-source software that enables humans to remotely control a robot for interacting with cloth and collecting demonstrations. Avigal et al. (2022) first annotated images with primitive actions and corresponding gripper positions and then trained a neural network to iteratively collect self-supervised data. Lee et al. (2021) required only 1 hour of random interactions with the cloth to develop their offline RL approach, which effectively handles complex sequential cloth folding (see Figure 3). In the new work of Lee et al. (2024), they further explored dataset collection by tackling the movement in human manipulation videos, making the dataset collection process more efficient and simpler.

Most works employ single-arm manipulation. Specifically, 65.9% (27 out of 41) of the reviewed papers use a single robotic arm, while 31.7% (13 out of 41) adopt a dual-arm system. Only one study (Xu et al., 2022) utilizes a three-arm setup. Regarding robot brands, the Universal Robots series (particularly UR5) is the most frequently used in real-world experiments, followed by the Franka Emika Panda arm. For dual-arm settings, existing solutions either (i) combine two independent single-arm robots into a coordinated dual-arm system (Ha and Song, 2022; Gu et al., 2024; Xue et al., 2023) or (ii) rely on dedicated dual-arm robot platforms such as ABB YuMi (Avigal et al., 2022), Kawada HIRO (Tanaka et al., 2021), or PR2 (Wu Y. et al., 2020).

The vision observation types include RGB, depth, RGB-D, grayscale, and point cloud images. We systematically analyze these inputs, which are used for training or inference, as shown in Table 3. RGB information is the most commonly used observation. Hoque et al. (2022a) compared the results of RGB, RGB-D, and depth, and suggested that RGB-D provides the best performance for their visual foresight task. Lin et al. (2022) demonstrated that RGB-related information are sensitive to camera views and visual features, which also poses challenges in the Sim2Real transformation. Later, Weng et al. (2022) utilized depth images as their policy input because they found that using depth images or point clouds as the training dataset could minimize the gap between simulation and reality.

The P&P configuration involves selecting a pick coordinate and a place coordinate. Initially, a robot grasps the cloth at the pick coordinate and lifts it to a certain height. The robot then moves above the place coordinate and finally places the cloth down and releases its grip. However, the P&P primitive action has its development. For example, Hoque et al. (2022a) provide pixel coordinates for the pick and place points, which the robot then uses to execute the action. Kase et al. (2022) decomposed and labeled the P&P action into finer phases: approach, grasp, pull, fold, release, and standby, to facilitate their cloth manipulation task. Avigal et al. (2022) incorporated the grasp angle into their manipulation process, estimating a pixel-wise value map for each gripper z-axis rotation to enhance the reliability of their P&P action. To achieve better performance, Blanco-Mulero et al. (2023) proposed a policy that optimizes parameters such as motion velocity and height within the P&P primitive action. Hietala et al. (2022) demonstrated that closed-loop feedback with parameterized P&P primitives significantly enhances adaptability in cloth manipulation, showing promise for more general and adaptive skills. Although the P&P action has been successfully utilized in cloth manipulation, it is relatively slow and constrained by the robot’s workspace.

To overcome the drawbacks of P&P actions, Ha and Song (2022) proposed a dynamic fling primitive that leverages object momentum for efficient unfolding. Their approach learns grasp locations while relying on predefined motion parameters, which may limit robustness across different cloth geometries or sizes. To address this issue, Gu et al. (2024) correlated the lift height with cloth size and adjusted the fling speed based on the height, which improved performance, as illustrated in Figure 4. Chen et al. (2022) further focused on learning fling action trajectories for garment unfolding using one robot arm, rather than a fixed fling trajectory.

Although the fling action can significantly expand the coverage, it is a coarse-grained manipulation method and is insufficient for fine-grained manipulation like P&P. Therefore, Canberk et al. (2023) strategically choose between the fling and P&P actions to efficiently and precisely unfold the garment. Their policy first utilizes the fling action to bring the cloth to a considerably unfolded state, then employs the P&P action to further expand it.

During cloth manipulation, predicted manipulation points may be outside the robot’s reachable workspace or correspond to difficult-to-handle cloth configurations when using previous primitive actions. Therefore, Avigal et al. (2022) and Xue et al. (2023) introduce a drag action in the unfolding stage. This action involves two robots simultaneously dragging the cloth away from its center. By exploiting the friction between the cloth and the supporting surface, the drag action helps smooth out corners and wrinkles, such as sleeves trapped underneath the garment, and can also be used to reposition the cloth. For the folding stage, Xue et al. (2023) further introduce a related action primitive mop. Similar to drag, mop is used to adjust the cloth position when the predicted grasp or placement points are unreachable during folding. Another form of dragging is described by He et al. (2023), in which two robotic arms are used asymmetrically: one arm grasps the cloth and remains stationary, while the other drags the cloth away by a predefined distance. This strategy is particularly effective for handling long sleeves that are covered by or folded inside a T-shirt.

The primitive actions described above manipulate the cloth either through sparse contact or by utilizing high speed robot. From the perspectives of dense force application and safety, Xu et al. (2022) proposed an air blow primitive action. In this approach, two robot arms grasp two points on the cloth, while a third arm operates a blower to apply air, expanding and unfolding the cloth. This action allows for the application of dense air forces on areas not in direct contact with the robot, thereby extending the system’s reach and enabling safe, high-speed interactions.

Early work on cloth manipulation often relies on explicit perception outputs, such as corners, edges, or segmentation masks, which are then mapped to hand-designed manipulation routines to realize simple folding or unfolding. Seita et al. (2019) used YOLO (Redmon et al., 2016) to detect blanket corners from depth images. The detected keypoints are passed to a keypoint-based heuristic controller: the robot grasps these points and pulls the blanket to increase coverage. Real-world experiments with two mobile manipulators and three blankets demonstrated strong generalization of this perception-guided heuristic pipeline. Several works adopt a similar structure for folding tasks, as illustrated in Figure 6. Canberk et al. (2023); He et al. (2023); Gu et al. (2024) use segmentation networks such as U-Net (Ronneberger et al., 2015), DeeplabV3 (Chen et al., 2017), and DeeplabV3+ (Chen et al., 2018) to localize keypoints or regions on flattened garments to fold the cloth.

While perception-guided heuristics are effective for predefined routines, they lack flexibility when target configurations vary. Goal-conditioned manipulation policies address this limitation by conditioning actions on both the current cloth state and a desired goal state, such that the predicted actions transform the cloth toward the goal. Weng et al. (2022) propose FabricFlowNet (FFN), a dual-arm goal-conditioned policy for cloth folding that leverages optical-flow prediction, (Figure 7). Instead of predicting actions directly from the current and goal images, FFN decomposes the policy into two components: a FlowNet that estimates particle flow between the current observation and the goal, and a PickNet that predicts P&P points from the estimated flow image. Mo et al. (2022) introduce Foldsformer, which incorporates space-time attention (Bertasius et al., 2021) into a folding planner. Given the current cloth image and a sequence of demonstration images, the model outputs a sequence of multi-step action points. This design balances speed and accuracy while capturing manipulation points and their ordering, even when cloth pose and size differ from those in the demonstration.

Goal-conditioned policies typically react to the current observation and goal, but do not explicitly model cloth dynamics or future evolution. Predictive and model-based methods aim to address this limitation by learning latent state representations or forward models of cloth behavior for planning and control. Hoque et al. (2022a) develop the VisuoSpatial Foresight (VSF) policy, trained on self-supervised simulated cloth manipulation data. VSF is built on Stochastic Variational Video Prediction (SV2P) (Babaeizadeh et al., 2018), an action-conditioned latent-variable video prediction model. At test time, the model receives the current and goal cloth states and predicts intermediate frames together with P&P coordinates, providing a visuospatial predictive model that can be used for planning. Ma et al. (2022) argue that human-defined labeled keypoints do not generalize well to unseen cloth configurations. They therefore use Transporter Networks (Kulkarni et al., 2019) to extract features and detect keypoints in an unsupervised manner from depth images. The detected keypoints are composed into a graph, and graph neural networks (GNNs) and recurrent networks are then used to model cloth dynamics in this learned space. Ganapathi et al. (2021) learn dense visual correspondences between different cloth configurations by training a Siamese network on pairs of cloth images, thereby capturing the underlying geometric structure. The learned correspondence field is used to transfer manipulation actions from a reference demonstration to new cloth states and has shown promising Sim2Real performance. Lin et al. (2022) propose a model-based RL approach in which a particle-based cloth dynamics model is learned from partial point clouds. A GNN models visible connectivity by operating on voxelized point clouds V and inferred edges E , and the learned dynamics are then used to train a P&P manipulation policy.

Rather than explicitly modeling cloth dynamics, reward-driven RL acquires manipulation strategies by optimizing task-specific reward functions through trial-and-error interaction. This paradigm learns the parameters of primitive actions (e.g., pick, drag, or fling) autonomously, but typically requires careful reward design and large amounts of interaction data. Wu Y. et al. (2020) introduce an RL framework for P&P cloth unfolding in which the placing policy is learned conditioned on random pick points. The final pick location is then chosen as the point with maximal value under the learned placing policy (the maximum-value-under-placing, MVP, strategy), leading to faster learning than jointly learning pick and place. Ha and Song (2022) employ Spatial Action Maps (Wu J. et al., 2020) for dynamic cloth manipulation. Their method evaluates a batch of candidate fling actions by transforming the observation and predicting a batch of value maps (Figure 8). The pixel with maximal value that also satisfies reachability constraints is selected, and its location and transformation are decoded into fling parameters. This value-map paradigm has been widely adopted in subsequent work on dynamic cloth manipulation, including He et al. (2023), Canberk et al. (2023), Gu et al. (2024), and Xu et al. (2022). Xu et al. (2022) improve the fling grasping strategy of Ha and Song (2022) by introducing edge-coincident grasp parameterizations to boost performance. Canberk et al. (2023) propose a factorized value-prediction model with two Spatial Action Maps networks (each with one encoder and two decoders) to generalize value maps over two primitive actions. Gu et al. (2024) apply this factorized policy to cloth unfolding and augment the value maps with an additional detection module. Lee et al. (2021) adopt an offline, batch-RL setting: a real robot first collects data via random actions, and a DQN is then trained on this fixed dataset, with DA used to improve robustness in low-data regimes. To scale reward-driven RL, Ha and Song (2022) further propose a self-supervised interaction framework in simulation: the robot interactively unfolds cloth, and the simulator computes coverage after each action. Episodes are reset once coverage reaches a threshold or an action limit is exceeded, eliminating the need for expert demonstrations or ground-truth state labels. Avigal et al. (2022) extend this idea to the real world by using a small set of human-labeled primitives and gripper poses for initial training, followed by large-scale self-supervised data collection.

Reward-driven RL can be sample-inefficient and sensitive to reward design, especially in real-world cloth manipulation. Demonstration-driven methods mitigate these challenges by leveraging expert demonstrations to provide more structured supervision. Seita et al. (2020) introduce behavior cloning (BC) for cloth unfolding. An Oracle supervisor generates unfolding demonstrations in simulation, and the policy is trained to imitate the supervisor from observed states. To improve robustness outside the demonstration distribution, they employ Dataset Aggregation (DAgger) (Ross et al., 2011), relabeling states visited under the learned policy, while DR over cloth appearance and camera poses supports Sim2Real transfer. Fu T. et al. (2023) propose a BC-based human-to-robot skill transfer framework for cloth unfolding. They decompose demonstrations into policy demonstrations (human-chosen P&P points) and action demonstrations (human manipulation trajectories), and use a mixture density network with parameter weighting to handle the multi-modal nature of unfolding behavior. The learned policy successfully unfolds cloth of various colors and sizes in the real world, with performance comparable to human operators. Lee et al. (2024) learn cloth manipulation actions directly from a small set of human videos (15 annotated demonstrations) to handle both unfolding and folding. A unified P&P policy is trained from these videos and deployed on a real robot, generalizing across fabrics with different shapes, colors, and textures. Other work, such as Goal-Aware GAIL (Tsurumine and Matsubara, 2022), explores adversarial imitation learning without hand-designed reward functions, but adversarial IL remains less common in current cloth manipulation studies.

While previous paradigms focus on learning low-level perception or control policies, they typically lack high-level semantic reasoning and task abstraction. LLM-based planning methods address this gap by leveraging large language or multimodal models to perform high-level decision making over manipulation primitives. Fu et al. (2024) introduces a large language model into cloth unfolding. They prompt ChatGPT with task requirements, a predefined taxonomy of cloth states, and corresponding operational primitives. The LLM then recommends which primitive to execute next. A segmentation network subsequently identifies manipulation points for the chosen primitive, combining LLM-based decision making with visual perception. Raval et al. (2024) first detects cloth corners using a perception module and converts them into structured representations, which, together with human instruction prompts, are provided to ChatGPT for high-level reasoning to determine P&P points for the robot. In addition, an unselected study, Deng et al. (2025), leverages vision-language models to predict manipulation plans from visual observations and semantic keypoints, whereas earlier approaches rely solely on text-based inputs.

Perception-guided heuristic methods are used for both unfolding and folding, particularly when some high-level geometric cues (e.g., corners, edges, contours) can be reliably extracted. By training detectors and segmentation networks on labeled cloth images (Seita et al., 2019; Gu et al., 2024), these approaches achieve high-accuracy perception and can directly localize manipulation-relevant regions on flattened or not severely wrinkled cloth, making them particularly suitable for simple unfolding and folding tasks. However, their reliance on hand-crafted post-processing and motion heuristics limits scalability. For example, downstream motion generation often ignores cloth deformability and contact dynamics, and heuristics tuned for one garment type or configuration frequently fail on heavily wrinkled, self-entangled, or topologically complex cloth, requiring manual redesign. In addition, the need for pixel-level labels and keypoints makes data collection expensive and restricts the diversity of cloth categories and configurations that can be covered. Future work could move from fixed geometric heuristics toward learned downstream controllers that take detector or segmentation outputs as input and are trained jointly with, or conditioned on, the follow-up actions. At the perception level, richer geometric and multimodal features (e.g., depth, 3D shape cues, or topological descriptors) and weaker forms of supervision could be used to reduce labeling costs. Extending these perception modules to operate robustly on non-flat, self-entangled garments and across a broader range of cloth categories would help perception-guided methods remain effective beyond narrowly structured folding scenarios.

Goal-conditioned manipulation policies frame cloth manipulation as a goal-reaching problem, mapping the current observation and a desired goal configuration to action sequences (Weng et al., 2022; Mo et al., 2022). This paradigm naturally aligns with folding tasks, whose target states are structured and can be expressed through goal images, keyframes, or demonstration trajectories. The primary challenge lies in goal specification and goal coverage. Existing approaches typically assume that a suitable goal image or trajectory is available and that the initial cloth state is not too far from this goal manifold. When the cloth is highly wrinkled, entangled, or heavily occluded, the system may struggle to infer a feasible goal-conditioned plan, and alternative paradigms (e.g., dynamics-based or RL-based methods) become more reliable. Future opportunities include allowing semantic goals (e.g., language descriptions, LLM-generated sub-goals) instead of explicit goal images, integrating predictive models to support longer-horizon reasoning, and learning goal manifolds that generalize across diverse garment categories and configurations. Such extensions would broaden the applicability of goal-conditioned policies beyond structured folding settings.

Predictive and model-based approaches focus on learning cloth dynamics or latent state representations that support downstream planning and control (Hoque et al., 2022a; Ma et al., 2022; Ganapathi et al., 2021; Lin et al., 2022). A key advantage of this paradigm is reusability. Once an accurate dynamics or representation model is learned, it can support multiple tasks, unfolding, flattening, or different folding patterns, simply by changing the planner or objective, without retraining the entire policy. This separation of representation and control also improves data efficiency, since costly robot interaction is used to fit the model once, and later tasks can rely on planning or offline optimization over the learned dynamics. However, model-based methods face several challenges. Long-horizon cloth dynamics are difficult to learn: small prediction biases accumulate quickly and can mislead planning. Video prediction and GNN-based models require large and diverse datasets, yet often struggle to represent task-critical properties such as layer ordering, self-occlusion, or contact conditions. Self-supervised objectives focused only on reconstruction or local geometry may also fail to encode physically meaningful structures. Promising directions include learning representations that better capture cloth topology and layer structure; integrating learned dynamics with model-based RL or planning under uncertainty; and incorporating multimodal cues, such as depth, force, or tactile feedback, to resolve ambiguities in partially observed states.

Reward-driven RL over primitive actions is currently the dominant paradigm for cloth unfolding (Wu Y. et al., 2020; Ha and Song, 2022; Canberk et al., 2023; Gu et al., 2024; Xu et al., 2022; Avigal et al., 2022). These methods optimize value or policy functions over discrete or continuous primitives using task-specific rewards. This paradigm is particularly well suited for cloth unfolding, because initial states are highly variable, heavily occluded, and partially observable. Such uncertainty naturally requires exploration and long-horizon decision-making, and RL agents can discover non-obvious sequences of pulls, flings, or drags that increase coverage even without an explicit cloth model. In practice, prior work has explored different ways to shape this learning process, for example by learning placement points from random picks (Wu Y. et al., 2020), designing multi-primitive policies that strategically select among several manipulation actions (Canberk et al., 2023), and tailoring reward functions to emphasize coverage (Ha and Song, 2022) or directional constraints in unfolding (Gu et al., 2024). Despite their strengths, RL methods face several challenges. Training in simulation is computationally expensive and strongly dependent on cloth and sensor fidelity, while large-scale real-world interaction is difficult to collect. Self-supervised interaction frameworks (Ha and Song, 2022; Avigal et al., 2022) reduce labeling effort, but significant Sim2Real gaps remain, and purely real-world training tends to be limited in scale and scenario diversity (Lee et al., 2021). Promising research directions include developing more realistic yet efficient cloth simulators, improving Sim2Real transfer methods, exploring more sample-efficient learning strategies, designing rewards that better capture cloth-specific manipulation objectives, and leveraging offline or data-driven RL methods capable of reusing large collections of prior trajectories.

Demonstration-driven methods learn manipulation skills directly from expert demonstrations using behavioral cloning or related imitation-learning techniques (Seita et al., 2020; Fu T. et al., 2023; Lee et al., 2024; Tsurumine and Matsubara, 2022). In practice, such methods have been applied to both cloth unfolding and folding. For unfolding, data-driven policies can clone expert sequences of pulls, flings, or shakes that gradually increase coverage. For folding and other structured subtasks, where expert strategies (e.g., aligning corners, placing creases, or executing a fixed folding sequence) are relatively consistent, demonstration-driven methods are particularly effective. Recent works further show that demonstrations collected in simulation, from real robots, or even from human videos can be transferred to robotic policies, sometimes with only a small number of annotated trajectories (Seita et al., 2020; Lee et al., 2021; 2024). However, these approaches are fundamentally limited by demonstration coverage and distribution shift. Policies often fail when encountering states that are not represented in the demonstrations, or when manipulating garments with different sizes, materials, and shapes. Thus, both the quality and the diversity of demonstrations are critical for robust performance across unfolding and folding scenarios. Future opportunities include improving data-collection platforms (Zhao et al., 2023; Wu et al., 2023) to make it easier to gather large, diverse, and high-quality demonstrations; combining offline imitation with selective online correction (e.g., DAgger-style relabeling) to mitigate distribution shift; and systematically evaluating how different types of demonstrations and input modalities (simulation rollouts, real-world robot teleoperation, and human-hand videos) influence generalization to new garments, initial configurations, and task variations.

LLM-based planning for cloth manipulation is still in its early stage. Fu et al. (2024); Raval et al. (2024); Deng et al. (2025) illustrate a transition from text-only LLM prompting to multimodal inputs that include cloth observations for guiding high-level cloth manipulation planning. Recent multimodal LLMs such as Gemini (Team et al., 2025) further suggest the feasibility of mapping visual observations directly to manipulation trajectories. Related progress in robotics, e.g., large-scale vision-language-action (VLA) models such as RT-2 (Zitkovich et al., 2023) and OpenVLA (Kim et al., 2024), as well as action-generation architectures such as Action Chunking Transformer (ACT) (Zhao et al., 2023) and Π 0.5 (Intelligence et al., 2025), which directly map images or language instructions to robot trajectories. Despite this promise, several challenges remain for cloth manipulation. First, collecting sufficiently diverse garment data at the scale required by LLMs and VLAs is difficult, and existing web-scale datasets contain little fine-grained deformable-object interaction. Second, inference latency and model size limit real-time deployment. Even models trained on specialized garment datasets are typically evaluated in quasi-static settings, where cloth deformation is slow and replanning frequency is low. Current action generators remain slower than conventional policy inference and often overlook cloth-specific dynamics such as self-occlusion, multilayer contact, and fast, large deformations, properties essential for dynamic interactions such as flinging, catching, or in-air regrasping. Promising directions therefore include finetuning or prompting LLMs and VLA models on cloth-manipulation datasets, distilling their plans into lightweight policies, and using them as high-level semantic planners that propose sub-goals or primitive sequences, while domain-specific controllers from the other paradigms (e.g., goal-conditioned policies, reward-driven RL, or demonstration-driven policies) execute those plans at a lower level and higher frequency. For action-generation architectures, a complementary avenue is to amortize expensive inference into compact latent plans or skill embeddings and let smaller, task-specific controllers decode short-horizon action chunks at control rate. This can both improve online speed and make such models more compatible with dynamic cloth primitives (e.g., fling or shake actions), enabling future systems to more fully exploit the flexible, highly deformable nature of garments rather than being limited to quasi-static operation.