Organizations: All the authors are with the Nara Institute of Science and Technology, Japan. · Christian Diehm is with the Karlsruhe Institute of Technology, Germany.
Robotic cloth untangling requires progressively disentangling fabric by adapting pulling actions to changing contact and tension conditions. Because large-scale real-world training is impractical due to cloth damage and hardware wear, sim-to-real policy transfer is a promising solution. However, cloth manipulation is highly sensitive to interaction dynamics, and policies that depend on precise force magnitudes often fail after transfer because similar force responses cannot be reproduced due to the reality gap. We observe that untangling is largely characterized by qualitative tension transitions rather than exact force values. This indicates that directly minimizing the sim-to-real gap in raw force measurements does not necessarily align with the task structure. We therefore hypothesize that emphasizing coarse force-change patterns while suppressing fine environment-dependent variations can improve robustness of sim-to-real transfer. Based on this insight, we propose Adaptive Force-Difference Quantization (ADQ), which reduces observation resolution by representing force inputs as discretized temporal differences and learning state-dependent quantization thresholds adaptively. This representation mitigates overfitting to environment-specific force characteristics and facilitates direct sim-to-real transfer. Experiments in both simulation and real-world cloth untangling demonstrate that ADQ achieves higher success rates and exhibits greater robustness in sim-to-real transfer than policies using raw force inputs. Supplementary video is available at https://youtu.be/ZeoBs-t0AWc
Figures & tables
Fig. 1: Overview of sim-to-real policy transfer for cloth untangling. In simulation, only the entangled portion of the cloth is modeled, and the policy is trained to progressively resolve the entanglement by fixing one hand (pin hand) and pulling with the other (pull hand). Then, the learned policy is transferred to the real world.
Fig. 2: Overview of ADQ. Raw force measurements are converted into force differences and discretized into three states (increase: +1 , decrease: −1 , no change: 0 ). The quantization thresholds are treated as policy-controlled variables, enabling adaptive sensitivity during execution. The resulting quantized force-difference representation is fed back into the policy to determine both the pulling actions and the thresholds. Policy training is performed within a reinforcement learning framework using this observation design.
Fig. 3: Local entanglement model for simulating cloth entangling. The entangled cloth portion is approximated as a chain of capsule segments connected by spherical joints where neighboring segments overlap to form a continuous body. To represent self-contact without introducing unstable persistent contacts, collisions are disabled for adjacent (overlapping) pairs and enabled only for non-neighboring pairs.
Fig. 4: Sim-to-Sim performance comparison in Gazebo. Lower (more negative) writhe reduction indicates better untangling. Statistical significance is assessed using one-sided Welch’s t-tests (Ours vs. each baseline; alternative: Ours is better), with Holm correction for multiple comparisons. * indicates significance level of p<0.05 . Each method was evaluated over n=10 trials.
Fig. 5: Snapshots of learned policy execution in the Gazebo simulator. The initial scene before policy execution and the final scene after execution are shown. An overview of the developed Gazebo-based cloth untangling environment is illustrated on the right side.
Fig. 6: Threshold sensitivity in Gazebo. We sweep the fixed ternary quantization threshold τ for ADQ w/o Adaptive and compare it with ADQ (Ours), which adapts τ online. Although ADQ w/o Adaptive was trained with τ=0.5 in Isaac Gym, its best-performing threshold in Gazebo shifts to τ=2.0 (average writhe reduction −7.3 ), indicating a hyperparameter shift across simulators. In contrast, ADQ (Ours) remains stable and achieves larger writhe reduction, reducing the manual tuning burden.
Randomized Variables
unit
min
max
Force-sensor scale coefficient
[-]
0.95
1.05
Force-sensor bias
N
−0.6
0.6
Pull-vector multiplier
[-]
0.05
10.0
Linear density of each chain
kg/m
0.267
0.445
Joint friction coefficient
[-]
0.375
0.625
Joint damping coefficient
[-]
0.375
0.625
TABLE I: Ranges of randomized simulation parameters
Fig. 7: Experimental setup showing the cloth/textile configurations used during sim-to-real testing.
Fig. 8: Snapshots of a tightly entangled case. The commanded pulling force is overlaid as a yellow arrow in each frame. The bottom plots show the observed force components (x,y,z) over time. Circled numbers denote the step index, and the step t at the bottom corresponds to the highlighted frame.
Configuration
Opposite
Naive
ADQ (Ours)
Loose
Low Friction
1.0
0.9
1.0
Medium Friction
1.0
0.9
1.0
High Friction
0.9
0.6
0.9
Tight
Low Friction
0.9
0.5
0.9
Medium Friction
1.0
0.5
1.0
High Friction
0.1
0.4
0.7
TABLE II: Success rate across materials and knot configurations under the sim-to-real protocol (success within 15 pulls and below a 30 N force safety limit). Each configuration was evaluated over n=10 trials.
Opposite
Naive
ADQ (Ours)
Configuration
pulls
peak
pulls
peak
pulls
peak
Loose
Low Friction
6.9
9.3
7.3
8.1
9.7
7.1
Medium Friction
7.4
10.3
7.7
8.5
11.1
6.9
High Friction
6.8
19.3
8.4
18.4
11.3
14.8
Tight
Low Friction
6.9
22.1
7.0
23.0
10.2
20.1
Medium Friction
7.0
23.0
7.6
19.8
9.3
17.2
TABLE III: Efficiency and safety statistics over successful trials. We report the average number of pulls (pulls) and the peak force (peak), defined as the maximum force observed during a pulling action, for each configuration.
Fig. 9: Analysis of policy behaviors. (a) Scene snapshots. The left arm is fixed, and the right arm pulls 3 cm per step. Yellow arrows denote the pulling direction. If the force exceeds 30 N, the gripper opens to ensure safety. The bottom three rows show different knot orientations and friction levels. (b) Grad-CAM features. Stacked areas indicate the feature contribution to the policy output. Solid lines show the pulling direction of x,y , and z . The pulling command ut is a normalized direction vector in [−1,1]3 . During execution, ut is scaled by α=3cm .
Fig. 10: Sim-to-real gap measured by the 1D Wasserstein distance between observation distributions (Isaac Gym: n=500 , Gazebo: n=150 ). Bars show bootstrap means with 95% CI (2000 resamples).
Although cloth is known to exhibit different outcomes under repeated fast dynamic motions, even when the same trajectory is applied, this variability has not yet been systematically characterized. Quantifying it is essential to assess the reliability of learned manipulation policies and the extent to which simulation can reproduce real-world behavior. To study this, we execute the same trajectory ten times across four dynamic tasks, two of which are novel, each tested with three cloths of very different properties and at up to three execution speeds, with a total of 269 recorded rollouts. For all of them, we record small marker positions on the cloth and synchronized stereo camera. We then formalize different metrics to quantify variability, and our results show how it is significant in every test condition, in most cases one to three orders of magnitude above the repeatability inherent to the robot and sensing noise. Our results also show variability is driven mainly by the cloth physical properties but also grows with speed. By replaying all the rollouts in four calibrated modern cloth simulators, we show that none of them can reproduce the variability magnitude we observed in real cloth, nor its ordering. Together with our analysis, we publish the dataset with synchronized OptiTrack, vision and robot logs and their corresponding simulator twins.
Mahed Dadgostar, Guillem Alenyà, Júlia Borràs
Institut de Robòtica i Informàtica Industrial, CSIC-UPC, Spain
Contact-rich manipulation requires robots to regulate both motion and interaction forces, yet achieving adaptive compliance remains a fundamental challenge. Learning from real-world data is costly and risky, while simulation-based approaches struggle with the sim-to-real gap in contact dynamics; existing sim-to-real methods either require real-world adaptation or sacrifice adaptive compliance by relying on isotropic compliant controllers. Our key insight is that force regulation decomposes into a time-varying but simulation-transferable directional component and a dynamics-sensitive but manually tunable magnitude component. We instantiate this directional component as two policy outputs, a task frame and a control mode vector, predicted by a visuomotor policy adapted from a pre-trained VLA model and trained via imitation learning on automatically generated simulation demonstrations. At deployment, an admittance controller integrates these predictions with human-specified stiffness and target wrench values to realize adaptive compliance. Our approach achieves adaptive compliance using only simulation data and can benefit from large-scale VLA pre-training. Extensive real-world experiments on four contact-rich tasks, microwave opening, peg-in-hole insertion, whiteboard wiping, and door opening, demonstrate strong task success rates and robustness to external disturbances. Project page: https://yifei-y.github.io/project-pages/TDC/.
Yifei Yang, Anzhe Chen, Zhenjie Zhu +6
Zhejiang University · Zhejiang Humanoid Robot Innovation Center
Linear-deformable manipulation remains challenging due to the complex deformations of objects such as cables and ropes. Prior data-driven approaches, particularly imitation learning, have shown some promise in narrowly defined settings but typically require thousands of demonstrations for specific tasks and cable types, limiting scalability and generalization. We introduce a sim-to-real reinforcement learning (RL) framework for multi-stage cable routing that leverages GPU-parallelized simulation to approximate linear deformable behaviors. Training across thousands of parallel simulations enables the learned policies to generalize across diverse cable geometries and deformation patterns. To bridge the sim-to-real gap, we propose a novel deployment strategy that combines a Simulation In the LOop (SILO) execution framework, localized RL policies, and robust cable state estimation. On real-world cable routing tasks, our approach achieves higher success rates and 2x reduction in cycle times compared to prior state-of-the-art learning methods. To our knowledge, this is the first successful sim-to-real transfer of RL policies for multi-stage cable routing. Videos and additional visualizations are available at https://silo-cable-routing.github.io/
Stone Tao, Jie Xu, Hesam Rabeti +3
1NVIDIA Corporation · University of California, San Diego