Go to ICLR 2026 Conference homepageDeformable Linear Object Manipulations with Differentiable Physics
Keywords: Deformable Linear Object Manipulation, Differentiable Physics
TL;DR: We introduce a differentiable simulation environment and benchmark for robotic learning in the manipulation of deformable linear objects.
Abstract: We address the challenge of enabling robots to manipulate deformable linear objects (DLOs), such as wires, ropes, and rubber bands. Prior work in this domain has primarily focused on narrow, task-specific problems, often relying on real-world demonstrations or handcrafted heuristics. Such approaches, however, do not scale to the diverse range of materials and tasks encountered in practice, where collecting sufficiently varied real-world data is impractical. Moreover, existing simulation environments provide limited support for the broad spectrum of material behaviors required for generalizable DLO manipulation. To overcome these limitations, we introduce a differentiable physics simulator specifically designed for versatile DLO manipulation. Our simulator models a wide range of material properties—including extensibility, inextensibility, elasticity, bending plasticity, and interactions with both rigid and deformable objects—thereby establishing a robust foundation for learning and evaluating manipulation skills. Building on this simulator, we propose a benchmark suite of representative DLO manipulation tasks that highlight their unique challenges. We further evaluate multiple policy learning algorithms on these tasks. The results show that reinforcement learning can learn closed-loop policies but requires prohibitively large amounts of data. In contrast, trajectory optimization is more efficient: gradient-based methods achieve the best sample efficiency when gradients are available, while sampling-based approaches are broadly applicable but less efficient.
Supplementary Material: zip
Primary Area: applications to robotics, autonomy, planning
Submission Number: 9432
Loading