Learning to unfold cloth: Scaling up world models to deformable object manipulation
1. Paper overview
In one sentence: Geometry-focused observations and demonstration-seeded, augmented replay make a DreamerV2 cloth controller more effective, while depth-sensor limits and garment-specific training constrain transfer. e02-taske05-modele06-normalse07-replaye08-traininge13-physicale14-limits
| At a glance | What to know |
|---|---|
| Research problem | Source description In-air unfolding requires reacting to changing three-dimensional folds and occluded grasp points. The task assumes the initial grasp and lift have succeeded: one corner is pinned, and a robot must grasp another suitable region and pull outward. Closed-loop control is intended to avoid rebuilding a complete geometric model after every disturbance. e02-task |
| Core mechanism | Source description Replace RGB observations with depth-derived surface-normal images to emphasize geometry and remove cloth color and texture from the visual representation. e06-normals |
| A key reported result | Zero-shot physical plane-cloth unfolding: Success: 72%, 68%, 76%, 80%, respectively; average 74%. Table III reports average timesteps of 107, 209, 212, 162, respectively, and an overall 172. Executed unfolding success; reported average timesteps. Franka Emika robot; 25 trials each for towel, dishcloth, facecloth and pillowcase; simulation-trained plane-cloth agent. No same-protocol physical baseline is tabulated. The authors explicitly caution that prior-work success rates are not directly comparable. Demonstrates bounded sim-to-real transfer. Failure includes timeout or an unsuccessful unfold without releasing the cloth. Timesteps cannot be converted to seconds without a control rate. e12-deploymente13-physical |
| Reading caution | Source description Near-contact wrist depth can fall inside the sensor minimum range, producing incomplete or noisy normals precisely during grasping. Curled or tangential corners are difficult to grasp; camera-layout and tactile alternatives are proposed, not quantitatively validated here. e12-deploymente14-limits |
Core contributions
- Source description
Replace RGB observations with depth-derived surface-normal images to emphasize geometry and remove cloth color and texture from the visual representation. e06-normals
- Source description
Seed world-model training with complete demonstration episodes and augment replayed image sequences consistently across time. These interventions address missing successful states and visual overfitting. e07-replay
- Source description
Evaluate four input/replay variants separately on five simulated garment types, then deploy the plane-cloth agent on a physical robot without additional training. e08-traininge10-garmentse12-deploymente13-physical
Figure 1. The three yellow additions change visual representation and the experience available to the world model. Original paper, p. 1 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin at the upper-right cameras and follow their arrows left into Depth2Normals. The converted images I join robot state S. During training, image/state/action/reward trajectories enter replay alongside demonstration episodes. The replay output splits: images pass through random augmentation, while state, action and reward feed the world model directly. A second path carries current images and state into the model. Follow latent Z_t to the RL policy, then action A_t back to the robot. This distinguishes the executed action loop from the stored training trajectories. The caption defines I, S, A, R and Z; the yellow blocks mark this paper’s additions. e03-interfacee05-modele07-replay
What it supports. The architecture makes the policy dependent on a learned representation of the current scene while using demonstrations and augmentation to improve model training. The source describes future-state and reward prediction inside the world-model approach. The visible action output comes from a separate policy block, supporting the report’s classical model-based RL classification.
Where the evidence stops. This is a system schematic, not a complete Dreamer computation graph. It omits actor-critic objectives and optimization paths; neither parameter sharing nor a separate online trajectory-search procedure can be inferred from the arrows.
2. Motivation
2.1 The problem and the proposed response
In-air unfolding requires reacting to changing three-dimensional folds and occluded grasp points. The task assumes the initial grasp and lift have succeeded: one corner is pinned, and a robot must grasp another suitable region and pull outward. Closed-loop control is intended to avoid rebuilding a complete geometric model after every disturbance. e02-task
2.2 What this reading follows
A hanging cloth changes shape as the robot reaches for it, so a useful controller must keep updating its understanding of folds and graspable regions. This paper adapts DreamerV2 around that requirement: two depth views become surface-normal images, demonstrations expose the world model to successful unfolding, and replay augmentation broadens the observations used for learning. The strongest evidence combines garment-wise simulation comparisons with a zero-shot physical plane-cloth experiment. Read those results at their actual scope: the initial lift is assumed, simulated garment types use separate agents, and the physical result covers four flat-cloth categories. e02-taske05-modele06-normalse07-replaye08-traininge13-physicale14-limits
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | Foundational work |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
This table preserves the labels recorded at reading time. The current major category is WAMs. View the current classification.
3.1 Evidence-based assessment
Supports the recorded classification
The recorded classical world-model/model-based RL category is supported by an action-conditioned dynamics/reward model feeding a separate RL policy. There is no described joint future-video/action generator or inverse-dynamics action extraction. The catalog’s architecture, prediction-paradigm and quadrant fields therefore remain not applicable to that generative taxonomy; shared training does not establish a One Model design. e05-model
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 1. Give the world model geometry that survives appearance changes
Imagine two episodes with similar cloth folds but different fabric colors. An RGB encoder must learn that the appearance change may be irrelevant to the next grasp. The paper instead computes surface-normal images from depth before learning, deliberately making local orientation prominent. Two viewpoints remain necessary in the described setup: the stand-off view supports scene-level positioning and the wrist view supports precise grasping. Robot position, gripper width and a tension flag supplement those images. The large Sobel kernel smooths orientation bands before reduction to 64×64. Reader interpretation: this shifts part of the representation problem into fixed preprocessing, which can simplify appearance transfer but also exposes the policy to the depth sensor’s failure modes. The reported near-contact failures show why geometry-focused input alone cannot guarantee robust grasping. e03-interfacee06-normalse12-deploymente14-limits
Figure 2. Depth is converted into a color-coded geometry representation before entering the learned model. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the central path from depth normalization to the output normal image. The x and y Sobel branches and the matrix of ones feed the normal-vector combination block. The subject-mask branch bypasses that combination and joins only at the final background-masking stage. Vector normalization and scaling into 0–255 occur before masking. These branch directions agree with the caption and method description. The caption specifies a Sobel kernel of size 9 and 256×256 processing; the surrounding text explains the final downsampling to the model’s 64×64 input. Broad, smooth color bands are intended to survive that reduction. e06-normalse12-deployment
What it supports. The image channels now encode surface orientation rather than the cloth’s visible color and texture. The authors intend wrinkles and creases to remain informative even when garment appearance changes. This is a deterministic observation transformation in the pipeline, used for both simulated and real camera depth before world-model inference.
Where the evidence stops. Figure 2’s caption mentions both 256×256 processing and 64×64 output; the text clarifies downsampling but leaves interpolation and masking settings unspecified. Geometry-focused images still inherit depth holes and near-contact sensor noise.
5.2 2. Make successful unfolding available before exploration discovers it
The paper’s replay argument starts with a data-coverage problem. A randomly exploring robot may rarely unfold the cloth, leaving its initial world model without observations of the goal configuration. Demonstration episodes supply complete task trajectories before training begins. When replay later samples a temporal image stack, augmentation applies consistently across that stack, preserving the relation between observations over time. The source excludes flips and excessive rotations because image directions must remain compatible with the control axes. These experiences train the model used by the separate RL policy; the paper does not describe direct behavior cloning of demonstration actions. Reader interpretation: demonstrations address which task states are seen, whereas augmentation addresses how those states can appear. Table II bundles both interventions, motivating a separate controlled test of each contribution. e05-modele07-replaye09-benchmarkse10-garments
5.3 3. Follow the evidence from simulated reward to physical success
The reward estimates cloth area from a coarse vertex grid, normalizes it by the initial maximum area and subtracts a timestep penalty. That training signal differs from both the normalized-return plot and the failure/near-success/success categories used in simulation evaluation. Keeping those quantities separate prevents a high curve from being mistaken for a success probability. Table II then tests garment-specific agents, revealing strong plane-cloth and short-sleeve results alongside shorts failures. The physical experiment takes only the plane-cloth agent into the real setup, where depth patching, synchronized sensing and an impedance controller connect the learned policy to actual motion. Its 74% average success therefore establishes a narrower result than universal garment manipulation. Reader interpretation: the most useful reproduction should test whether the proposed observation and replay mechanisms explain this bounded transfer, rather than only matching one aggregate score. e04-rewarde08-traininge10-garmentse11-curvese12-deploymente13-physical
5.4 Training and inference
During training
Train each input/replay configuration for one million simulated timesteps per garment type: twenty agents total. Randomize cloth position, rotation, size, color and texture, plus the end-effector starting position. The paper does not establish a held-out garment-identity split. e08-training
The heuristic demonstration script knows the target corner and cloth dimensions and can regrasp. Replay augmentations may change brightness, contrast, hue, saturation or value and apply slight rotation, translation or zoom; flips and excessive rotations are excluded to preserve control-axis consistency. e07-replaye09-benchmarks
During inference
Synchronize stereo-camera images and proprioception, convert coordinates to Unity conventions, patch depth occlusions and compute normals. The pretrained agent supplies displacements to a real-time Cartesian impedance controller. New observations close the feedback loop; the source does not specify a separate online trajectory-search algorithm. e05-modele12-deployment
Physical deployment is zero-shot for the plane-cloth policy. Trials vary cloth properties, placement, pinned corner and initial fold condition. Learned predictions support control, while success is assessed from executed robot manipulation. e12-deploymente13-physical
5.5 Implementation flow
- Construct observations and actions
Unity, RFUniverse and ObiCloth provide the simulated task. The four action components are scaled to [-1,1]. The two cameras provide complementary scene and grasp views; robot-state features accompany the images. e03-interfacee06-normals
- Convert depth to geometry-focused images
Normalize depth, apply x/y Sobel operators, combine their outputs with a matrix of ones, normalize the resulting vectors, scale into 0–255 and mask the background. Figure 2 specifies a size-9 Sobel kernel and 256×256 processing; the model receives downsampled 64×64 images. e06-normals
- Learn dynamics and a policy
Demonstration and collected image/state/action/reward trajectories enter replay. Sampled images pass through augmentation before world-model training. The paper describes initial world-model learning followed by actor-critic learning after environment interaction. Figure 1 routes latent state to a distinct RL policy, whose actions return to the robot. e05-modele07-replay
6. Experiments & results
This DreamerV2 adaptation learns to unfold cloth already suspended from one corner. Depth-derived surface normals, demonstration-seeded replay and temporally consistent image augmentation improve several simulated garment results. A plane-cloth policy transfers zero-shot to physical trials with 74% average success; this is not a universal garment policy. The method and deployment evidence are detailed in e02-task, e05-model, e07-replay and e13-physical.
6.1 Read the original evidence
Figure 5. The learning curves show how quickly useful unfolding behavior emerges under the stated training setup. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the legend to distinguish modified Dreamer in blue from vanilla Dreamer in orange, then compare their rise across the training budget. Section IV.A specifies one million timesteps, which clarifies the horizontal scale. The vertical axis is normalized return: episode reward is divided by the maximum episode timesteps and shifted upward by one. It is not percent success. The caption states that image inputs are surface normals and that early stopping occurs at 80% unfold. Vector SAC instead receives cloth-vertex coordinates. The green RAIF line corresponds to the R-AIF comparison described in the text, using the same demonstration episodes as the modified model. e08-traininge09-benchmarkse11-curves
What it supports. The modified Dreamer trend improves earlier and stays above vanilla Dreamer in the displayed experiment. R-AIF also improves substantially, whereas the SAC curves remain low. This supports the paper’s qualitative learning-efficiency argument under its chosen setup, but the image alone does not establish a statistically reliable ordering across repeated training runs.
Where the evidence stops. Smoothed lines and pale fluctuations are not defined confidence intervals. Smoothing settings and seed variability are absent, and training data or privileged observations differ across baselines. Do not read endpoint heights as garment success rates or wall-clock speed.
Table III. A simulation-trained plane-cloth policy succeeds in 74% of the reported physical trials. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the success column across the four physical cloth categories, then the average row. Each category has 25 trials. The accompanying text says cloth properties, placement, pinned corner and fold condition vary between trials. These are deployments of the plane-cloth agent, not physical tests of all five simulated garment-specific agents. The table’s retained note defines failure as exceeding the designated timesteps or performing an unsuccessful unfold without releasing the cloth. The final column reports average timesteps, including the printed overall value of 172; it does not report seconds. Section IV.B supplies the camera, coordinate-conversion and impedance-control context needed to interpret the experiment. e02-taske12-deploymente13-physicale14-limits
What it supports. The displayed successes—72%, 68%, 76% and 80%—support zero-shot transfer for these flat-cloth categories under the corner-suspended setup. The result concerns real robot execution, not visual predictions. It provides a practical demonstration of transfer while leaving a substantial fraction of trials unsuccessful and broader garment transfer untested.
Where the evidence stops. No same-protocol real-world baseline or uncertainty interval accompanies this table. Control frequency and the designated timeout are not specified. The paper warns that published success rates from other systems are not directly comparable to these trials.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Zero-shot physical plane-cloth unfolding Franka Emika robot; 25 trials each for towel, dishcloth, facecloth and pillowcase; simulation-trained plane-cloth agent. | Success: 72%, 68%, 76%, 80%, respectively; average 74%. Table III reports average timesteps of 107, 209, 212, 162, respectively, and an overall 172. Executed unfolding success; reported average timesteps | No same-protocol physical baseline is tabulated. The authors explicitly caution that prior-work success rates are not directly comparable. Demonstrates bounded sim-to-real transfer. Failure includes timeout or an unsuccessful unfold without releasing the cloth. Timesteps cannot be converted to seconds without a control rate. e12-deploymente13-physical |
| Simulated garment unfolding after one million training timesteps Separate agent per garment/configuration; 250 evaluation trials per agent; episode classification uses maximum achieved unfold. | Normals with modified replay: long-sleeve 72.0%, short-sleeve 100%, plane-cloth 84.8%, shorts 17.6%, unbuttoned shirt 42.4%. Success percentage, Table II | Pick/place: 64.2%, 64.0%, 28.8%, 55.2%, 3.2%, in the same garment order. On plane-cloth, RGB vanilla, RGB modified and normals vanilla score 2.0%, 19.6% and 36.0%. The complete combination improves four garment categories over the script, but loses substantially on shorts. Its plane-cloth gain does not isolate demonstrations from augmentation, because both change with replay configuration. e08-traininge09-benchmarkse10-garments |
| Simulation learning-curve benchmark Figure 5 uses surface normals for image-input agents; vector SAC receives a 4×4 cloth-vertex grid. R-AIF uses the same demonstrations as the modified model. | Qualitatively, the modified Dreamer curve rises earlier and remains above vanilla Dreamer in the displayed run; no exact endpoint is transcribed. Smoothed normalized episode return | Curves include vanilla Dreamer, image SAC, vector SAC, scripted motion and R-AIF. Return divides episode rewards by maximum episode timesteps and adds one. Early stopping occurs at 80% unfold. This is not a success-percentage curve, and the plot supplies no defined uncertainty band. e09-benchmarkse11-curves |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table II. The combined method wins on four garment groups; shorts expose a clear exception. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each garment group as three percentages in the caption’s order: failure, near success, success. Compare down a group to hold garment type fixed. The four learned rows cross RGB or normals with vanilla or modified replay, while the last row is the scripted baseline. Training uses a separate agent for each garment/configuration and one million timesteps; the caption gives 250 evaluation trials per agent. In the plane-cloth group, success progresses from 2.0% for RGB vanilla to 19.6% for RGB modified, 36.0% for normals vanilla and 84.8% for normals with modified replay. The final row provides the 28.8% pick/place comparison. e07-replaye08-traininge09-benchmarkse10-garmentse15-reporting
What it supports. The strongest combination reaches 100% short-sleeve success and 84.8% plane-cloth success in this evaluation. Yet it achieves only 17.6% on shorts, below the script’s 55.2%. These garment-wise differences are more informative than a blanket claim of superiority and support treating the method as specialized controllers with uneven task performance.
Where the evidence stops. Modified replay changes both demonstration initialization and augmentation, so this table cannot isolate their individual effects. The source gives no uncertainty estimates or numeric boundaries between failure, near success and success; the script also has target-corner information.
7. Analysis & limitations
7.1 What the evidence leaves open
Near-contact wrist depth can fall inside the sensor minimum range, producing incomplete or noisy normals precisely during grasping. Curled or tangential corners are difficult to grasp; camera-layout and tactile alternatives are proposed, not quantitatively validated here. e12-deploymente14-limits
Small pixel footprint is the authors’ explanation for shorts failures, not an isolated causal result. Large garments were excluded by the single-arm setup, and each garment type required its own agent. e10-garmentse14-limits
The bundled replay ablation cannot establish that demonstrations alone are essential. Table II lacks uncertainty estimates and precise fail/near/success thresholds. The prose’s aggregate gain claims are not transparently reconciled with the garment-wise table; retain individual results. e07-replaye10-garmentse15-reporting
7.2 Questions for discussion
- Would separating demonstration seeding from image augmentation preserve the plane-cloth improvement?
- Do higher-resolution observations improve shorts while holding task geometry and replay data fixed?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction needs the Unity/RFUniverse/ObiCloth task, demonstration generator, dual-view geometry pipeline and DreamerV2 configuration. The PDF omits demonstration count, model/actor-critic loss details and weights, optimizer settings, latent size, imagination horizon, training hardware, package versions and randomization ranges. e03-interfacee05-modele07-replaye08-training
Deployment also needs camera calibration and depth-hole patching, coordinate conversion, tension-flag acquisition and impedance-control settings. Camera models, control frequency, timeout, strain threshold and complete success criteria are unspecified; agree on these before comparing physical success. e03-interfacee04-rewarde12-deploymente13-physical
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Separate demonstration coverage from replay augmentation
Reader-proposed check, not performed: on plane-cloth with surface-normal inputs fixed, train four conditions crossing demonstration initialization on/off with sequence-consistent augmentation on/off. Use equal initial replay sizes, equal environment-interaction and optimization budgets, matched randomization schedules and repeated seeds. Compare both normalized-return learning curves and final success under explicitly fixed evaluation thresholds. Track time to the first successful episode and success on a common set of perturbed camera poses. If demonstrations improve early learning but augmentation improves camera-offset robustness, their distinct roles are supported. If either isolated intervention matches the full combination across both outcomes, the claim that both are needed would weaken. Establish missing demonstration counts and hyperparameters before treating this as an exact reproduction. e07-replaye08-traininge09-benchmarkse10-garmentse14-limits
Check 2: Test whether shorts failures are caused by limited visual resolution
Reader-proposed check, not performed: compare the original 64×64 observation pipeline with higher-resolution normals and a garment-centered crop resized to the original input size. Retrain matched shorts agents using the same replay interventions, demonstrations, physical garment sizes, action limits and evaluation initial states. Include a plane-cloth control and the scripted baseline; record grasp-location errors, failure/near/success outcomes, model size and inference cost. The crop condition helps distinguish pixel coverage from a larger network alone. Improvement concentrated on previously tiny or poorly localized corners would support the authors’ pixel-footprint explanation. Persistent shorts failure despite better localization would instead point toward garment geometry or manipulation strategy. Any crop-tracking failures must remain in the evaluation. e03-interfacee06-normalse08-traininge09-benchmarkse10-garmentse14-limits
8.3 Reading coverage
Visual audit: Visually inspected all eight PDF pages, including the title/authors/version on p. 1, task scope on p. 2, interface and reward equations on p. 3, geometry pipeline on p. 4, replay/training/evaluation protocol on p. 5, quantitative and diagnostic material on p. 6, deployment limitations on p. 7 and references on p. 8. Figures 1–5 and Tables I–III were inspected; five final original crops were also individually viewed. Figure 1’s data-flow arrows and Figure 2’s mask/Sobel branches were checked against their captions and method text. Table notes needed to interpret outcome order, trial counts and failure definitions are retained. All method, numeric and reproduction claims are supported on these declared pages. No appendix is present; separate supplemental material remains unverified, and code and experiments are outside this reading.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8. Appendix coverage: not present.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Abstract; I. Introduction (pp. 1–2)
- II. Related Work (pp. 2–3)
- III.A. States, Actions, and Reward Function (p. 3)
- III.B. Model, including surface normals and replay-buffer modifications (pp. 3–5)
- IV.A. Simulation: Benchmarking and Garment Manipulation (pp. 5–6)
- IV.B. Real-world evaluation (pp. 6–7)
- V. Conclusions and limitations (p. 7)
- References (p. 8)
Outside the original text pass
- The inspected title page identifies arXiv:2602.16675v1, 18 February 2026. Title and all three authors match the supplied catalog; no revision discrepancy was observed. Other versions were not supplied or compared.
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- The extraction limitation was addressed by visually inspecting all eight PDF pages and all final figure/table crops.
- Separate supplemental material availability has not been fully verified.
- No separate supplement was supplied. No appendix is present in this eight-page PDF. Code was not inspected and experiments were not reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01-identityPDF p. 1, title/author block and arXiv margin stamp
Title matches the supplied catalog. Authors are Jack Rome (University of Edinburgh), Stephen James (Imperial College London), and Subramanian Ramamoorthy (University of Edinburgh). The stamp reads arXiv:2602.16675v1 [cs.RO], 18 Feb 2026.
Go to primary source ↓e02-taskPDF pp. 1–2, I. Introduction and II. Related Work, paragraph beginning ‘In this work, we start’
The problem is closed-loop in-air unfolding after a successful corner grasp and lift. Initial pickup is explicitly outside the task.
Go to primary source ↓e03-interfacePDF p. 3, III.A, first three paragraphs
Unity with RFUniverse and ObiCloth simulates corner-pinned cloth. Actions control XYZ displacement and gripper state. Observations are two 64×64×3 images, XYZ gripper position, width and tension boolean.
Go to primary source ↓e04-rewardPDF p. 3, III.A, Eqs. (1)–(2) and adjacent explanation
Reward uses triangle areas from a 4×4 vertex grid, normalized by pre-gravity spawn area, minus t. Prose describes a penalty subtracting 1 and a strain constraint without providing its threshold.
Go to primary source ↓e05-modelPDF p. 1, Figure 1 and caption; pp. 3–4, III.B. Model
Image/state/action/reward replay trains a world model; images are augmented separately. Current images/state enter the model, latent Z_t feeds RL Policy, and action A_t feeds the robot. Text describes future-state/reward prediction and later actor-critic learning but does not provide network/loss configurations.
Go to primary source ↓e06-normalsPDF p. 4, Figure 2 and caption; Figure 3; III.B.1 and preceding camera paragraph
Depth normalization branches into x/y Sobel filters, a ones matrix and a subject mask. Vector normalization and 0–255 scaling precede background masking. Caption specifies kernel 9 and 256×256 processing; text describes downsampling to 64×64 and complementary stand-off/wrist views.
Go to primary source ↓e07-replayPDF pp. 4–5, III.B.2. Modifying the replay buffer; p. 1, Figure 1
Demonstrations replace random initial episodes and expose successful states. Augmentations apply uniformly to sampled image stacks; listed transformations exclude flipping or excessive rotation. Demonstration and augmentation effects are bundled.
Go to primary source ↓e08-trainingPDF p. 5, IV.A.2. Garment Manipulation
Four configurations cross RGB/normals with vanilla/modified replay. Each of five garment types has separately trained agents, twenty total, each trained for 1M timesteps. Cloth and robot-start variations are randomized; outcomes use highest achieved unfold.
Go to primary source ↓e09-benchmarksPDF p. 5, IV.A.1. Benchmarking
The heuristic has target-corner and dimension knowledge and can regrasp. SAC uses either the image observation space or 4×4 vertex coordinates. R-AIF uses the same demonstration episodes. Normalized return divides rewards by maximum episode timesteps and adds one.
Go to primary source ↓e10-garmentsPDF p. 6, Table II, all garment groups and rows; caption
Cells are fail/near/success percentages over 250 trials per agent. Modified normals success is 72.0/100/84.8/17.6/42.4 for long-sleeve/short-sleeve/plane-cloth/shorts/unbuttoned shirt. Pick/place success is 64.2/64.0/28.8/55.2/3.2. Plane-cloth RGB vanilla/RGB modified/normals vanilla success is 2.0/19.6/36.0.
Go to primary source ↓e11-curvesPDF p. 6, Figure 5, axes, legend and caption
Plot shows smoothed normalized returns against training steps for six methods including RAIF. Caption specifies surface-normal image inputs and 80% unfold early stopping. Exact final values, uncertainty definitions and smoothing parameters are not tabulated.
Go to primary source ↓e12-deploymentPDF pp. 6–7, IV.B. Real-world evaluation
Franka Emika uses stand-off and wrist stereo cameras, synchronized observations, Unity-coordinate conversion, patched depth and Cartesian impedance control. The plane-cloth agent is deployed zero-shot. Sensor minimum range and noisy near-contact observations hinder grasping.
Go to primary source ↓e13-physicalPDF p. 6, Table III and caption; p. 7, IV.B opening paragraph
Twenty-five trials per category give towel/dishcloth/facecloth/pillowcase success of 72/68/76/80%, average 74%; corresponding average timesteps are 107/209/212/162, printed overall average 172. Failure covers timeout or failed unfold without release. Prior-work comparisons are explicitly not direct.
Go to primary source ↓e14-limitsPDF p. 7, IV.B final paragraphs and V. Conclusions
Authors discuss curled/tangential corners, wrist-depth range failure, limited pixels for shorts, exclusion of large garments by single-arm constraints, and separately trained policies per garment. Alternative sensing, larger observations and dual arms are future directions.
Go to primary source ↓e15-reportingPDF p. 5, IV.A.2 results paragraph; p. 6, Table II and Figure 4
Prose claims an average 15% success increase and 35% failure reduction without a transparent aggregation calculation. Table II gives individual garment percentages; Figure 4 supplies illustrative outcome examples, not numeric classification thresholds.
Go to primary source ↓8.5 Primary sources
Learning to unfold cloth: Scaling up world models to deformable object manipulation ↗
PDF · 6,343 extracted words
Source fingerprint
bd88e04642c648e26f19e4921e0ae8010c8fec3d28619a15320f0ec96776531d