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

Jul 16, 2026

Upload by Maeve

Linear-deformable manipulation remains difficult because cables and ropes undergo complex, local deformations driven by contact, gravity, and accumulated slack. Prior imitation learning methods often need thousands of demonstrations per cable type and task, limiting scalability. This work presents a sim-to-real reinforcement learning framework for multi-stage cable routing that uses GPU-parallelized simulation to approximate linear-deformable dynamics, enabling policy generalization. A SILO deployment strategy with localized RL policies and robust cable state estimation improves real-world success rates and reduces cycle times by about 2× versus prior methods.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages28

Reading Time1 h 11 min

Introduction

Problem of Linear-Deformable Cable Manipulation

Learning Reactive Policies via Sim-to-Real RL

System Overview

GPU-Parallelized Simulation of Linear Deformables

Localized RL Policies and Motion Primitives

SILO Deployment and State Estimation

arXiv :2607 .046 16v 1 [ cs .RO] 6 Jul 2026 SILO: Simulation-in-the-Loop Sim-to-Real Transfer for Multi-Stage Cable Routing Stone Tao 1 ,2 , Jie Xu 1 , Hesam Rabeti 1 , Yashraj Narang 1 , Yijie Guo 1∗ , Iretiayo Akinola 1∗ 1NVIDIA Corporation, 2University of California, San Diego 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 simto-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 ∼2× 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/](https://silo-cable-routing.github.io/) Keywords: Sim-to-Real, Cable Routing, Reinforcement Learning 1 Introduction Industrial automation remains a central objective in robotics, yet its success has largely been confined to rigid body assembly where dynamics are predictable and execution can be governed by scripted trajectories or perception-driven motion planning pipelines. However, many real-world assembly processes, such as automotive wiring, hose routing, and textile handling require manipulating deformable objects. The high variability and dimensionality of deformable materials demand continuous, dexterous low-level control, posing a challenge for traditional robotic frameworks. In this paper, we address the cable manipulation problem by learning reactive low-level policies with reinforcement learning (RL) in simulation followed by sim-to-real deployment. We focus on cable routing, a ubiquitous industrial task that requires handling subtle, local deformations during execution. These deformations are difficult to predict from visual observations alone and can change abruptly due to contact, gravity, and accumulated slack. As a result, cable routing requires policies that can react to sudden changes in cable deformations. Traditional systems often require substantial task-specific engineering to operate reliably in tightly constrained settings. Moreover, the unpredictable nature of deformable objects necessitates frequent replanning, making real-time, high-performance execution difficult to achieve. As an alternative to traditional planning based systems, data-driven methods present a more general approach to train robotics policies. Imitation learning based approaches [1, 2] leverage task-specific data, and in some cases large pre-training datasets are used in VLAs [3, 4, 5] . While imitation learning is fairly general, it requires sufficient high-quality demonstrations that can be costly or impractical to collect for deformable object tasks. The work most related to ours is the hierarchical imitation learning approach [6] that demonstrated promising results for multi-stage cable routing, however it requires 1000s of demonstrations. In contrast, RL-based sim-to-real approaches require Figure 1: Key components of our system. A) GPU-parallelized articulated rigid-body simulation approximates cable dynamics during routing. B) Motion primitives grasp and transport the simulated cable t

FAQs

Why is multi-stage cable routing difficult for traditional robotics approaches?

Cables deform locally and can change abruptly due to contact, gravity, and accumulated slack, making reliable low-level control and real-time replanning challenging.

What does the proposed method use to train policies before deployment?

It trains reactive low-level policies in simulation with reinforcement learning, using GPU-parallelized simulation to approximate linear-deformable cable behaviors.

How does the approach bridge the sim-to-real gap during real robot execution?

It deploys a SILO (Simulation In the LOop) framework that synchronizes the robot in simulation and real execution, combined with localized RL policies and robust cable state estimation to reduce controller dynamics mismatch and avoid harness collisions.

Jul 16, 2026

Upload by Maeve