Pith. sign in

REVIEW 2 major objections 67 references

RL policies trained on rigid-link cable approximations transfer zero-shot to multi-stage real-world cable routing when the robot tracks simulated joint targets in a loop.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-11 16:21 UTC pith:6A335IEH

load-bearing objection Solid systems paper: first real-robot multi-stage cable-routing RL via rigid-capsule sim + SILO, with clean ablations; the headline vs h-IL over-attributes gains to method rather than known harness poses. the 2 major comments →

arxiv 2607.04616 v1 pith:6A335IEH submitted 2026-07-06 cs.RO cs.AI

SILO: Simulation-in-the-Loop Sim-to-Real Transfer for Multi-Stage Cable Routing

Stone Tao , Jie Xu , Hesam Rabeti , Yashraj Narang , Yijie Guo , Iretiayo Akinola This is my paper
classification cs.RO cs.AI
keywords sim-to-realcable routingreinforcement learninglinear deformable manipulationSimulation-In-the-LoopGPU-parallelized simulationstate estimation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Routing a cable through a sequence of harnesses is hard because small local deformations change suddenly under contact and gravity, so classical planners and imitation-learning systems either over-engineer or demand thousands of demonstrations. This paper shows that a coarse rigid-body model—an articulated chain of capsules with PD-driven joints plus a single plasticity term—can be simulated thousands of times in parallel on a GPU, letting reinforcement learning discover reactive routing policies for a single harness. At deployment a Simulation-In-the-Loop (SILO) controller keeps a digital twin synchronized with the real robot: the policy’s action is applied only inside the twin, and the real arm simply tracks the resulting joint angles. Combined with a polyline state estimator that matches the simulated cable discretization, the same policy routes four different real cables through up to three harnesses with higher success and roughly half the cycle time of the previous hierarchical imitation baseline, without any demonstrations or system identification.

Core claim

A GPU-parallelized rigid-body approximation of a linear deformable, localized RL that only learns the critical routing step, and a Simulation-In-the-Loop deployment loop that re-uses the same simulator dynamics at runtime together constitute the first successful zero-shot sim-to-real transfer of RL policies for multi-stage cable routing, yielding higher success rates and approximately 2× lower cycle times than prior learning methods while generalizing across cable materials.

What carries the argument

SILO (Simulation-In-the-Loop): at every control step the real joint angles are copied into a digital twin, the policy acts only inside that twin, and the resulting simulated joint targets become the real robot’s set-points; this re-uses training dynamics, bypasses controller system-ID, and enforces collision-free targets by construction.

Load-bearing premise

A chain of rigid capsules whose joints are driven by fixed PD gains and one plasticity coefficient is already close enough to real cable behavior that synchronizing joint angles alone closes the sim-to-real gap.

What would settle it

Train the identical policy and SILO pipeline, then measure real-world success on a cable whose bending stiffness or diameter lies far outside the range of the four tested materials (or on harness clearances smaller than the cable diameter); a sharp drop below the reported 14–18/24 rates would falsify the approximation claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 0 minor

Summary. The paper presents SILO, a sim-to-real RL system for multi-stage cable routing. Linear deformables are approximated as articulated rigid-capsule chains with PD elasticity and a plasticity coefficient β, trained with PPO on GPU-parallelized intermediate states generated by motion primitives. At deployment, SILO (Algorithm 1) applies policy actions only in a digital twin and tracks the resulting simulated joint targets on the real robot, combined with polyline state estimation from stereo depth and SAM2. Real-robot experiments report higher success rates and roughly 2× lower cycle times than a hierarchical imitation-learning baseline across 1–3 harnesses, with generalization to four cable types and ablations on randomization, deployment method, controller gains, and observation dimensionality.

Significance. If the results hold under fairer controls, this is a concrete advance for industrial deformable assembly: the first demonstrated zero-shot sim-to-real RL transfer for multi-stage cable routing, with real hardware success rates (Tables 1–2), controlled ablations (Tables 3–5, 11), cycle-time breakdown, and emergent reactive behaviors (angling/swinging). The rigid-capsule approximation plus SILO controller-agnostic deployment is a practical engineering contribution that avoids system identification and simplifies reward design. Strengths include reproducible real-robot numbers, explicit failure-mode taxonomy, and open videos. The work is significant for robotics venues even if the headline baseline comparison is imperfect, because the SILO stack and localization strategy are independently useful.

major comments (2)
  1. Table 1 and §5.1: The central claim of higher success (18/24 vs 12/24 for H3) and ~2× lower cycle time vs hierarchical IL is load-bearing for the abstract and introduction, yet the comparison confounds SILO/RL with privileged knowledge. SILO assumes known harness geometries/poses a priori (Problem Description, Limitations) and uses MoveToHarness to place the TCP at a known p_i before localized RL; h-IL does not receive that prior and spends time on reshaping/replanning. The authors note the comparison is imperfect, but still headline numerical superiority and “first successful.” A matched-assumption baseline (same known fixtures + same recovery budget) or an ablation that removes the known-pose prior is needed before attributing the gain primarily to SILO + rigid-capsule RL.
  2. §4.1 and Appendix E: The cable model (fixed PD stiffness/damping + single plasticity β, Table 8) is an ad-hoc free-parameter stack frozen once for all cables. Table 2 shows generalization across four real cables, which is encouraging, but there is no sensitivity study or identification procedure showing how β and joint gains map to material properties, nor whether the same parameters would transfer to substantially different lengths, diameters, or contact regimes. Without that, the claim that the approximation is “sufficient for zero-shot transfer” remains under-supported for broader linear-deformable tasks.

Circularity Check

0 steps flagged

No circularity: empirical robotics system whose success rates and cycle times are measured on held-out real hardware, not derived by construction from fitted inputs or self-citations.

full rationale

The paper presents an engineering system (GPU rigid-capsule cable approximation + localized PPO + SILO joint-target tracking + polyline state estimation) whose central claims are zero-shot real-robot success rates (Table 1: 24/24, 22/24, 18/24) and cycle times (~87 s) versus an external hierarchical-IL baseline. Hyper-parameters (joint stiffness/damping, plasticity β, PD gains, reward coefficients) are chosen once and frozen; they do not appear inside any equation that is later re-labeled a “prediction.” SILO (Algorithm 1) simply re-uses the same simulator controller that generated the training data; it does not embed the success metric. Self-citations (ManiSkill3 infrastructure) supply only the parallel rigid-body backend and are not load-bearing uniqueness or uniqueness theorems. No step reduces a claimed first-principles result or prediction to its own inputs by definition. The non-apples-to-apples baseline comparison noted by the skeptic is a fairness/correctness issue, not circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central empirical claim rests on a small set of modeling choices (rigid-capsule cable + fixed PD/plasticity, known harness poses, quasi-static SILO sync) and a handful of free coefficients that are never derived from first principles. No new physical entities are postulated beyond the engineering construct “SILO.”

free parameters (5)
  • cable joint drive stiffness = 1e-2
    Set to 1e-2 (Table 8); controls elastic restoring force and is not measured from real cables.
  • cable joint drive damping = 1e-3
    Set to 1e-3; free damping coefficient of the PD joint drives.
  • plasticity coefficient β = 0.7
    Set to 0.7; governs how quickly rest angles track current angles; chosen by hand for “sufficient state coverage.”
  • cable Y-axis randomization range = [-0.2, 0.2] m
    Training distribution width [-0.2, 0.2] m that directly controls entry-angle coverage; ablated but still a free design choice.
  • simulation PD gains (low/medium/high) = medium (default)
    Three discrete controller settings used for SILO flexibility experiments; medium is default; none are identified from real hardware under SILO.
axioms (4)
  • domain assumption Harness geometries and poses are known a priori (from CAD or a separate perception pipeline).
    Stated in §3 and Limitations; without it the digital twin and MoveToHarness primitive cannot be constructed.
  • ad hoc to paper Linear deformables can be approximated for the routing task by an articulated sequence of rigid capsules with 3 orthogonal revolute joints plus PD elasticity and simple plasticity.
    Core modeling choice of §4.1; no derivation from continuum mechanics is offered.
  • domain assumption Manipulation is quasi-static so that joint-angle synchronization between real robot and twin at 10 Hz is sufficient.
    Explicit assumption of SILO (Appendix C.1) and listed as a limitation for dynamic tasks.
  • ad hoc to paper Four equal-length cable points nearest the TCP plus TCP pose and joint angles form a sufficient observation for reactive routing.
    Observation design of §4.3; ablated but still an unproven sufficiency claim.
invented entities (2)
  • SILO (Simulation-In-the-Loop) deployment loop no independent evidence
    purpose: Bypass controller system identification by executing policy actions only inside a synchronized digital twin and commanding the real robot to track the resulting joint targets.
    Novel engineering construct introduced in Algorithm 1 and §4.3; no independent theoretical existence outside this paper.
  • Plasticity-augmented rigid-capsule cable model no independent evidence
    purpose: Provide a GPU-parallel, high-throughput surrogate for elastic/plastic linear deformables that still yields transferable policies.
    Constructed in §4.1 from existing PhysX APIs; parameters are free and not independently validated against continuum models.

pith-pipeline@v1.1.0-grok45 · 27662 in / 3356 out tokens · 30783 ms · 2026-07-11T16:21:28.955491+00:00 · methodology

0 comments
read the original abstract

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/

Figures

Figures reproduced from arXiv: 2607.04616 by Hesam Rabeti, Iretiayo Akinola, Jie Xu, Stone Tao, Yashraj Narang, Yijie Guo.

Figure 1
Figure 1. Figure 1: Key components of our system. A) GPU-parallelized articulated rigid-body simulation approxi￾mates cable dynamics during routing. B) Motion primitives grasp and transport the simulated cable to ran￾domized intermediate states near the harness to efficiently train localized RL policies. C) At deployment, the robot executes identical motion primitives before transitioning to simulation-in-the-loop execution. … view at source ↗
Figure 3
Figure 3. Figure 3: Three images showing A) the real world, B) rendering of the digital twin, C) an overlay of the task [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: Four example intermediate states of our cable simulation generated by mo￾tion primitives for initializing episodes dur￾ing RL training. The simulated Franka robot gripper is grasping the end of a cable and is placed just behind a harness at pose pi. Our key design choice is to model linear-deformables us￾ing only rigid-body APIs, trading off fidelity for high￾throughput simulation to enable fast RL trainin… view at source ↗
Figure 4
Figure 4. Figure 4: Observation features in simulation. The red dots denote the 4 observed cable points closest to the TCP. The RGB axes show the observed TCP pose. In simulation, cable points are obtained directly from the link states, where each link has equal length. In the real world, in order to estimate sim-to-real com￾patible cable points, we fit an equal-length polyline to a segmented cable pointcloud. To predict the … view at source ↗
Figure 5
Figure 5. Figure 5: Sampling distribution of harness poses for cable routing. 1 to 3 harnesses are independently sampled within one of three 12.5cm×15cm regions, with in-plane rotations uniformly sampled up to 45◦ . The green boxes are spaced out by 5 cm. We evaluate on a cable routing benchmark consisting of routing a single cable through one to three harnesses. We follow the experimental protocol of Luo et al. [6], which us… view at source ↗
Figure 6
Figure 6. Figure 6: Reset states used to densely eval￾uate trained policies. The cable is shown already routed through the target harness (top) to show what a solved state looks like. To better understand what enables our method to bridge the sim-to-real dynamics gap, we conduct ablation studies on our simulation environment. To keep tests controlled and fast, we analyze the performance of our policy in rout￾ing through a sin… view at source ↗
Figure 7
Figure 7. Figure 7: Visual overview of the state estimation system used to build sim digital twins during deployment. In [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The sequence of movements for the GraspCable primitive in the real world [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: The sequence of movements for the GraspCable primitive in simulation. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The sequence of movements for the MoveToHarness primitive in the real world [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: The sequence of movements for the MoveToHarness primitive in simulation. B.1 GraspCable Visualization of the GraspCable primitive deployed in the real world is shown in [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Selected frames of two overlaid videos showing how using SILO with a digital twin that models the [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Comparison of code complexity between SILO and a typical sim-to-real deployment setup. This [PITH_FULL_IMAGE:figures/full_fig_p020_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Motion profile of each of the 7 arm joints after calibrating the simulation controller to align with the [PITH_FULL_IMAGE:figures/full_fig_p023_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Training success rate over environment samples. Curve is the average of 5 trials, and shaded area is [PITH_FULL_IMAGE:figures/full_fig_p025_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Image of the real world setup for experiments. [PITH_FULL_IMAGE:figures/full_fig_p025_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Cantilever test comparing the 4 main cables tested in this paper. [PITH_FULL_IMAGE:figures/full_fig_p026_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Selected frames of the angling behavior demonstrated by the RL policy when routing through a [PITH_FULL_IMAGE:figures/full_fig_p026_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Selected frames of the swinging behavior demonstrated by the RL policy when routing through [PITH_FULL_IMAGE:figures/full_fig_p026_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Two common failure types during execution where the RL policy fails to route completely. Left [PITH_FULL_IMAGE:figures/full_fig_p028_20.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

67 extracted references · 8 canonical work pages

  1. [1]

    C. Chi, S. Feng, Y . Du, Z. Xu, E. Cousineau, B. Burchfiel, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion. InProceedings of Robotics: Science and Systems (RSS), 2023

    2023

  2. [2]

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn. Learning fine-grained bimanual manipulation with low-cost hardware. In K. E. Bekris, K. Hauser, S. L. Herbert, and J. Yu, editors,Robotics: Science and Systems XIX, Daegu, Republic of Korea, July 10-14, 2023, 2023. doi:10.15607/ RSS.2023.XIX.016. URLhttps://doi.org/10.15607/RSS.2023.XIX.016

    doi:10.15607/rss.2023.xix.016 2023

  3. [3]

    Black, N

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, et al.π0: A vision-language-action flow model for general robot control. corr, abs/2410.24164, 2024. doi: 10.48550.arXiv preprint ARXIV .2410.24164

    Pith/arXiv arXiv 2024

  4. [4]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. P. Foster, P. R. Sanketi, Q. Vuong, T. Kollar, B. Burchfiel, R. Tedrake, D. Sadigh, S. Levine, P. Liang, and C. Finn. Openvla: An open-source vision-language-action model. In P. Agrawal, O. Kroemer, and W. Burgard, editors,Proceedings of The 8th Conference on Robot Lea...

    2025

  5. [5]

    Bjorck, N

    NVIDIA, J. Bjorck, N. C. Fernando Casta ˜neda, X. Da, R. Ding, L. J. Fan, Y . Fang, D. Fox, F. Hu, S. Huang, J. Jang, Z. Jiang, J. Kautz, K. Kundalia, L. Lao, Z. Li, Z. Lin, K. Lin, G. Liu, E. Llontop, L. Magne, A. Mandlekar, A. Narayan, S. Nasiriany, S. Reed, Y . L. Tan, G. Wang, Z. Wang, J. Wang, Q. Wang, J. Xiang, Y . Xie, Y . Xu, Z. Xu, S. Ye, Z. Yu, ...

    2025

  6. [6]

    J. Luo, C. Xu, X. Geng, G. Feng, K. Fang, L. Tan, S. Schaal, and S. Levine. Multi-stage cable routing through hierarchical imitation learning.arXiv pre-print, 2023. URLhttps: //arxiv.org/abs/2307.08927

    Pith/arXiv arXiv 2023

  7. [7]

    Noseworthy, B

    M. Noseworthy, B. Tang, B. Wen, A. Handa, C. C. Kessens, N. Roy, D. Fox, F. Ramos, Y . Narang, and I. Akinola. FORGE: force-guided exploration for robust contact-rich ma- nipulation under uncertainty.IEEE Robotics Autom. Lett., 10(5):4436–4443, 2025. doi: 10.1109/LRA.2025.3551637. URLhttps://doi.org/10.1109/LRA.2025.3551637

    work page doi:10.1109/lra.2025.3551637 2025

  8. [8]

    B. Tang, M. A. Lin, I. Akinola, A. Handa, G. S. Sukhatme, F. Ramos, D. Fox, and Y . Narang. Industreal: Transferring contact-rich assembly tasks from simulation to reality. In K. E. Bekris, K. Hauser, S. L. Herbert, and J. Yu, editors,Robotics: Science and Systems XIX, Daegu, Republic of Korea, July 10-14, 2023, 2023. doi:10.15607/RSS.2023.XIX.039. URLhtt...

    work page doi:10.15607/rss.2023.xix.039 2023

  9. [9]

    T. He, W. Xiao, T. Lin, Z. Luo, Z. Xu, Z. Jiang, J. Kautz, C. Liu, G. Shi, X. Wang, L. J. Fan, and Y . Zhu. HOVER: versatile neural whole-body controller for humanoid robots. InIEEE International Conference on Robotics and Automation, ICRA 2025, Atlanta, GA, USA, May 19-23, 2025, pages 9989–9996. IEEE, 2025. doi:10.1109/ICRA55743.2025.11128549. URL https:...

    doi:10.1109/icra55743.2025.11128549 2025

  10. [10]

    Rudin, D

    N. Rudin, D. Hoeller, P. Reist, and M. Hutter. Learning to walk in minutes using mas- sively parallel deep reinforcement learning. In A. Faust, D. Hsu, and G. Neumann, ed- itors,Conference on Robot Learning, 8-11 November 2021, London, UK, volume 164 of Proceedings of Machine Learning Research, pages 91–100. PMLR, 2021. URLhttps: //proceedings.mlr.press/v...

    2021

  11. [11]

    Azulay, K

    O. Azulay, K. Kondap, J. Drake, S. Xie, H. Li, S. Chitta, and K. Goldberg. MOTORCYCLE 1.0: Automating bimanual cable routing around fixtures on the NIST task board. In21st IEEE 10 International Conference on Automation Science and Engineering, CASE 2025, Los Angeles, CA, USA, August 17-21, 2025, pages 2636–2641. IEEE, 2025. doi:10.1109/CASE58245.2025. 111...

    doi:10.1109/case58245.2025 2025

  12. [12]

    K. Chen, Z. Bing, F. Wu, Y . Meng, A. Kraft, S. Haddadin, and A. Knoll. Contact-aware shaping and maintenance of deformable linear objects with fixtures. InIROS, pages 1–8, 2023. doi:10.1109/IROS55552.2023.10341726. URLhttps://doi.org/10.1109/IROS55552. 2023.10341726

    doi:10.1109/iros55552.2023.10341726 2023

  13. [13]

    S. Jin, W. Lian, C. Wang, M. Tomizuka, and S. Schaal. Robotic cable routing with spatial representation.IEEE Robotics Autom. Lett., 7(2):5687–5694, 2022. doi:10.1109/LRA.2022. 3158377. URLhttps://doi.org/10.1109/LRA.2022.3158377

    doi:10.1109/lra.2022 2022

  14. [14]

    G. A. Waltersson, R. Laezza, and Y . Karayiannidis. Planning and control for cable-routing with dual-arm robot. In2022 International Conference on Robotics and Automation, ICRA 2022, Philadelphia, PA, USA, May 23-27, 2022, pages 1046–1052. IEEE, 2022. doi: 10.1109/ICRA46639.2022.9811765. URLhttps://doi.org/10.1109/ICRA46639.2022. 9811765

    doi:10.1109/icra46639.2022.9811765 2022

  15. [15]

    Z. Chen, Y . Chitambar, T. Lam, H. Li, S. Chitta, and K. Goldberg. Craft: Long-horizon cable routing algorithm and low-friction caging gripper. InProceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2026

    2026

  16. [16]

    S. Dong, S. Wang, Y . She, N. Sunil, A. Rodriguez, and E. H. Adelson. Cable manipulation with a tactile-reactive gripper. In M. Toussaint, A. Bicchi, and T. Hermans, editors,Robotics: Science and Systems XVI, Virtual Event / Corvalis, Oregon, USA, July 12-16, 2020, 2020. doi: 10.15607/RSS.2020.XVI.029. URLhttps://doi.org/10.15607/RSS.2020.XVI.029

    work page doi:10.15607/rss.2020.xvi.029 2020

  17. [17]

    M. Li, H. Yu, Y . Huang, Y . Hong, and C. Choi. Hierarchical DLO routing with reinforcement learning and in-context vision-language models.CoRR, abs/2510.19268, 2025. doi:10.48550/ ARXIV .2510.19268. URLhttps://doi.org/10.48550/arXiv.2510.19268

    Pith/arXiv arXiv doi:10.48550/arxiv.2510.19268 2025

  18. [18]

    Schulman, A

    J. Schulman, A. X. Lee, J. Ho, and P. Abbeel. Tracking deformable objects with point clouds. In2013 IEEE International Conference on Robotics and Automation, Karlsruhe, Germany, May 6-10, 2013, pages 1130–1137. IEEE, 2013. doi:10.1109/ICRA.2013.6630714. URL https://doi.org/10.1109/ICRA.2013.6630714

    work page doi:10.1109/icra.2013.6630714 2013

  19. [19]

    M. Yan, Y . Zhu, N. Jin, and J. Bohg. Self-supervised learning of state estimation for ma- nipulating deformable linear objects.IEEE Robotics Autom. Lett., 5(2):2372–2379, 2020. doi:10.1109/LRA.2020.2969931. URLhttps://doi.org/10.1109/LRA.2020.2969931

    doi:10.1109/lra.2020.2969931 2020

  20. [20]

    S. Jin, D. Romeres, A. Ragunathan, D. K. Jha, and M. Tomizuka. Trajectory optimization for manipulation of deformable objects: Assembly of belt drive units. InIEEE International Conference on Robotics and Automation, ICRA 2021, Xi’an, China, May 30 - June 5, 2021, pages 10002–10008. IEEE, 2021. doi:10.1109/ICRA48506.2021.9561556. URLhttps:// doi.org/10.11...

    doi:10.1109/icra48506.2021.9561556 2021

  21. [21]

    H. Kim, Y . Ohmura, and Y . Kuniyoshi. Multi-task real-robot data with gaze attention for dual-arm fine manipulation. InIEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2024, Abu Dhabi, United Arab Emirates, October 14-18, 2024, pages 8516–

    2024

  22. [22]

    doi:10.1109/IROS58592.2024.10802034

    IEEE, 2024. doi:10.1109/IROS58592.2024.10802034. URLhttps://doi.org/10. 1109/IROS58592.2024.10802034

    doi:10.1109/iros58592.2024.10802034 2024

  23. [23]

    Hoang, H

    T. Hoang, H. Le, P. Becker, N. A. Vien, and G. Neumann. Geometry-aware RL for manipula- tion of varying shapes and deformable objects. InThe Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025. OpenReview.net, 2025. URLhttps://openreview.net/forum?id=7BLXhmWvwF. 11

    2025

  24. [24]

    C. Li, R. Zhang, J. Wong, C. Gokmen, S. Srivastava, R. Mart ´ın-Mart´ın, C. Wang, G. Levine, M. Lingelbach, J. Sun, M. Anvari, M. Hwang, M. Sharma, A. Aydin, D. Bansal, S. Hunter, K. Kim, A. Lou, C. R. Matthews, I. Villa-Renteria, J. H. Tang, C. Tang, F. Xia, S. Savarese, H. Gweon, C. K. Liu, J. Wu, and L. Fei-Fei. BEHA VIOR-1K: A benchmark for embodied A...

    2022

  25. [25]

    URLhttps://proceedings.mlr.press/v205/li23a.html

  26. [26]

    X. Lin, Y . Wang, J. Olkin, and D. Held. Softgym: Benchmarking deep reinforcement learning for deformable object manipulation. In J. Kober, F. Ramos, and C. J. Tomlin, editors,4th Conference on Robot Learning, CoRL 2020, 16-18 November 2020, Virtual Event / Cambridge, MA, USA, volume 155 ofProceedings of Machine Learning Research, pages 432–448. PMLR,

    2020

  27. [27]

    URLhttps://proceedings.mlr.press/v155/lin21a.html

  28. [28]

    E. Xing, V . Luk, and J. Oh. Stabilizing reinforcement learning in differentiable multiphysics simulation. InThe Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025. OpenReview.net, 2025. URLhttps://openreview. net/forum?id=DRiLWb8bJg

    2025

  29. [29]

    Z. Weng, P. Zhou, H. Yin, A. Kravberg, A. Varava, D. Navarro-Alarcon, and D. Kragic. In- teractive perception for deformable object manipulation.IEEE Robotics Autom. Lett., 9(9): 7763–7770, 2024. doi:10.1109/LRA.2024.3431943. URLhttps://doi.org/10.1109/ LRA.2024.3431943

    work page doi:10.1109/lra.2024.3431943 2024

  30. [30]

    Matas, S

    J. Matas, S. James, and A. J. Davison. Sim-to-real reinforcement learning for deformable object manipulation. In2nd Annual Conference on Robot Learning, CoRL 2018, Z ¨urich, Switzerland, 29-31 October 2018, Proceedings, volume 87 ofProceedings of Machine Learn- ing Research, pages 734–743. PMLR, 2018. URLhttp://proceedings.mlr.press/v87/ matas18a.html

    2018

  31. [31]

    P. M. Scheikl, E. Tagliabue, B. Gyenes, M. Wagner, D. Dall’Alba, P. Fiorini, and F. Mathis- Ullrich. Sim-to-real transfer for visual reinforcement learning of deformable object manip- ulation for robot-assisted surgery.IEEE Robotics Autom. Lett., 8(2):560–567, 2023. doi: 10.1109/LRA.2022.3227873. URLhttps://doi.org/10.1109/LRA.2022.3227873

    doi:10.1109/lra.2022.3227873 2023

  32. [32]

    Y . Wu, W. Yan, T. Kurutach, L. Pinto, and P. Abbeel. Learning to manipulate deformable objects without demonstrations. In M. Toussaint, A. Bicchi, and T. Hermans, editors,Robotics: Science and Systems XVI, Virtual Event / Corvalis, Oregon, USA, July 12-16, 2020, 2020. doi: 10.15607/RSS.2020.XVI.065. URLhttps://doi.org/10.15607/RSS.2020.XVI.065

    work page doi:10.15607/rss.2020.xvi.065 2020

  33. [33]

    Coumans and Y

    E. Coumans and Y . Bai. Pybullet, a python module for physics simulation for games, robotics and machine learning.http://pybullet.org, 2016–2019

    2016

  34. [34]

    Isaac Sim

    NVIDIA. Isaac Sim. URLhttps://github.com/isaac-sim/IsaacSim

  35. [35]

    Todorov, T

    E. Todorov, T. Erez, and Y . Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026–

    2012

  36. [36]

    doi:10.1109/IROS.2012.6386109

    IEEE, 2012. doi:10.1109/IROS.2012.6386109

    doi:10.1109/iros.2012.6386109 2012

  37. [37]

    Newton: GPU-accelerated physics simulation for robotics, and simula- tion research., 2025

    Newton Contributors. Newton: GPU-accelerated physics simulation for robotics, and simula- tion research., 2025. URLhttps://github.com/newton-physics/newton

    2025

  38. [38]

    Xiang, Y

    F. Xiang, Y . Qin, K. Mo, Y . Xia, H. Zhu, F. Liu, M. Liu, H. Jiang, Y . Yuan, H. Wang, L. Yi, A. X. Chang, L. J. Guibas, and H. Su. SAPIEN: A simulated part-based inter- active environment. In2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 11094–11104. Computer Vision Foundation / ...

    doi:10.1109/cvpr42600.2020.01111 2020

  39. [39]

    J. Gu, F. Xiang, X. Li, Z. Ling, X. Liu, T. Mu, Y . Tang, S. Tao, X. Wei, Y . Yao, X. Yuan, P. Xie, Z. Huang, R. Chen, and H. Su. Maniskill2: A unified benchmark for generalizable manipulation skills. InInternational Conference on Learning Representations, 2023

    2023

  40. [40]

    S. Tao, F. Xiang, A. Shukla, Y . Qin, X. Hinrichsen, X. Yuan, C. Bao, X. Lin, Y . Liu, T. kai Chan, Y . Gao, X. Li, T. Mu, N. Xiao, A. Gurha, V . N. Rajesh, Y . W. Choi, Y .-R. Chen, Z. Huang, R. Calandra, R. Chen, S. Luo, and H. Su. Maniskill3: Gpu parallelized robotics simulation and rendering for generalizable embodied ai.Robotics: Science and Systems, 2025

    2025

  41. [41]

    M. Macklin. Warp: A high-performance python framework for gpu simulation and graphics. https://github.com/nvidia/warp, March 2022. NVIDIA GPU Technology Conference (GTC)

    2022

  42. [42]

    Huang, Y

    Z. Huang, Y . Hu, T. Du, S. Zhou, H. Su, J. B. Tenenbaum, and C. Gan. Plasticinelab: A soft- body manipulation benchmark with differentiable physics. In9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenRe- view.net, 2021. URLhttps://openreview.net/forum?id=xCcdBRQEDW

    2021

  43. [43]

    Hu, T.-M

    Y . Hu, T.-M. Li, L. Anderson, J. Ragan-Kelley, and F. Durand. Taichi: a language for high- performance computation on spatially sparse data structures.ACM Transactions on Graphics (TOG), 38(6):201, 2019

    2019

  44. [44]

    M. Li, R. Antonova, D. Sadigh, and J. Bohg. Learning tool morphology for contact-rich manip- ulation tasks with differentiable simulation. InIEEE International Conference on Robotics and Automation, ICRA 2023, London, UK, May 29 - June 2, 2023, pages 1859–1865. IEEE, 2023. doi:10.1109/ICRA48891.2023.10161453. URLhttps://doi.org/10.1109/ICRA48891. 2023.10161453

    doi:10.1109/icra48891.2023.10161453 2023

  45. [45]

    J. Xu, T. Chen, L. Zlokapa, M. Foshey, W. Matusik, S. Sueda, and P. Agrawal. An End-to- End Differentiable Framework for Contact-Aware Robot Design. InProceedings of Robotics: Science and Systems, Virtual, July 2021. doi:10.15607/RSS.2021.XVII.008

    work page doi:10.15607/rss.2021.xvii.008 2021

  46. [46]

    Jacob, T

    J. Jacob, T. Bandyopadhyay, J. Williams, P. V . K. Borges, and F. Ramos. Learning to simulate tree-branch dynamics for manipulation.IEEE Robotics Autom. Lett., 9(2):1748–1755, 2024. doi:10.1109/LRA.2024.3349830. URLhttps://doi.org/10.1109/LRA.2024.3349830

    work page doi:10.1109/lra.2024.3349830 2024

  47. [47]

    Makoviychuk, L

    V . Makoviychuk, L. Wawrzyniak, Y . Guo, M. Lu, K. Storey, M. Macklin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, and G. State. Isaac gym: High performance gpu-based physics simula- tion for robot learning, 2021

    2021

  48. [48]

    Jiang, C

    Y . Jiang, C. Wang, R. Zhang, J. Wu, and L. Fei-Fei. Transic: Sim-to-real policy transfer by learning from online correction. InConference on Robot Learning, 2024

    2024

  49. [49]

    Christen, W

    S. Christen, W. Yang, C. P ´erez-D’Arpino, O. Hilliges, D. Fox, and Y .-W. Chao. Learning human-to-robot handovers from point clouds. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  50. [50]

    Y . Qin, B. Huang, Z. Yin, H. Su, and X. Wang. Dexpoint: Generalizable point cloud reinforce- ment learning for sim-to-real dexterous manipulation. In K. Liu, D. Kulic, and J. Ichnowski, editors,Conference on Robot Learning, CoRL 2022, 14-18 December 2022, Auckland, New Zealand, volume 205 ofProceedings of Machine Learning Research, pages 594–605. PMLR,

    2022

  51. [51]

    URLhttps://proceedings.mlr.press/v205/qin23a.html

  52. [52]

    Tobin, R

    J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In2017 IEEE/RSJ Inter- national Conference on Intelligent Robots and Systems, IROS 2017, V ancouver , BC, Canada, 13 September 24-28, 2017, pages 23–30. IEEE, 2017. doi:10.1109/IROS.2017.8202133....

    doi:10.1109/iros.2017.8202133 2017

  53. [53]

    Singh, A

    R. Singh, A. Allshire, A. Handa, N. D. Ratliff, and K. V . Wyk. Dextrah-rgb: Visuomotor policies to grasp anything with dexterous hands.CoRR, abs/2412.01791, 2024. doi:10.48550/ ARXIV .2412.01791. URLhttps://doi.org/10.48550/arXiv.2412.01791

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.01791 2024

  54. [54]

    Zakka, B

    K. Zakka, B. Tabanpour, Q. Liao, M. Haiderbhai, S. Holt, J. Y . Luo, A. Allshire, E. Frey, K. Sreenath, L. A. Kahrs, C. Sferrazza, Y . Tassa, and P. Abbeel. Mujoco playground: An open- source framework for gpu-accelerated robot learning and sim-to-real transfer., 2025. URL https://github.com/google-deepmind/mujoco_playground

    2025

  55. [55]

    Mittal, P

    M. Mittal, P. Roth, J. Tigue, A. Richard, O. Zhang, P. Du, A. Serrano-Mu ˜noz, X. Yao, R. Zurbr ¨ugg, N. Rudin, L. Wawrzyniak, M. Rakhsha, A. Denzler, E. Heiden, A. Borovicka, O. Ahmed, I. Akinola, A. Anwar, M. T. Carlson, J. Y . Feng, A. Garg, R. Gasoto, L. Gulich, Y . Guo, M. Gussert, A. Hansen, M. Kulkarni, C. Li, W. Liu, V . Makoviychuk, G. Malczyk, H...

    Pith/arXiv arXiv 2025

  56. [56]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms.CoRR, abs/1707.06347, 2017. URLhttp://arxiv.org/abs/1707.06347

    Pith/arXiv arXiv 2017

  57. [57]

    B. Wen, M. Trepte, J. Aribido, J. Kautz, O. Gallo, and S. Birchfield. Foundationstereo: Zero- shot stereo matching.arXiv, 2025

    2025

  58. [58]

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. R ¨adle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V . Alwala, N. Carion, C.-Y . Wu, R. Girshick, P. Dol- lar, and C. Feichtenhofer. SAM 2: Segment anything in images and videos. InThe Thirteenth International Conference on Learning Representations, 2025. URLhttps://openreview. net...

    2025

  59. [59]

    B. Wen, S. Dewan, and S. Birchfield. Fast-foundationstereo: Real-time zero-shot stereo match- ing.arXiv preprint arXiv:2512.11130, 2025

    arXiv 2025

  60. [60]

    Carion, L

    N. Carion, L. Gustafson, Y .-T. Hu, S. Debnath, R. Hu, D. Suris, C. Ryali, K. V . Alwala, H. Khedr, A. Huang, et al. Sam 3: Segment anything with concepts.arXiv preprint arXiv:2511.16719, 2025

    Pith/arXiv arXiv 2025

  61. [61]

    Schneider

    T. Schneider. franky: High-Level Control Library for Franka Robots. URLhttps://github. com/TimSchneider42/franky

  62. [62]

    Berscheid and T

    L. Berscheid and T. Kr ¨oger. Jerk-limited real-time trajectory generation with arbitrary target states.Robotics: Science and Systems XVII, 2021

    2021

  63. [63]

    T. Lin, K. Sachdev, L. Fan, J. Malik, and Y . Zhu. Sim-to-real reinforcement learning for vision-based dexterous manipulation on humanoids.arXiv:2502.20396, 2025. 14 For videos, see the supplemental video / our website: https://silo-cable-routing.github.io/ A State Estimation System Figure 7 shows the pipeline used to predict cable points and construct th...

    Pith/arXiv arXiv 2025

  64. [64]

    This provides a 1D ordering parameter along the cable length

    Cable Pointcloud Projection: We compute the dominant directionσof the point cloud using singular value decomposition (SVD) and project all points onto this axis. This provides a 1D ordering parameter along the cable length

  65. [65]

    Points falling within each bin are averaged to produce an initial ordered polyline

    Binning: The projected points are normalized to[0,1]×[0,1]×[0,1]and partitioned into devenly spaced bins, wheredis the cable decimation. Points falling within each bin are averaged to produce an initial ordered polyline. If no points fall into a bin (e.g., due to occlusion), the closest valid polyline point is reused to maintain continuity

  66. [66]

    This procedure is robust to moderate occlusions (e.g., partial harness blockage), as disconnected visible segments are implicitly reconnected during polyline construction

    Resampling: The polyline is resampled to enforce equal spacing between adjacent points, ensuring compatibility with the fixed-length link representation used in simulation. This procedure is robust to moderate occlusions (e.g., partial harness blockage), as disconnected visible segments are implicitly reconnected during polyline construction. The method i...

  67. [67]

    plasticity

    uses a state estimation system for cables but takes a different approach. Their approach can handle more turns and handle some crossovers in cables compared to our system, but otherwise is limited to overhead cameras and planar cables. In contrast, our system predicts the 3D positions of points on the cable and is thus not limited to overhead cameras or p...

    2015