An Overview of Simulators for Deformable Object Manipulation
(last updated March 11th, 2026)
Introduction
Physics simulation has become an indispensable tool in robotics research, providing controlled, repeatable and cost effective environments for developing and evaluating manipulation strategies prior to real-world deployment. While the simulation of rigid bodies is now relatively mature with well established contact models, stable solvers and manageable sim-to-real transfer, the simulation of deformable objects remains a substantially harder and largely open problem.
Deformable objects encompass a broad range of materials encountered in everyday robotic tasks: soft volumetric bodies such as foam, rubber and biological tissue; deformable linear objects such as cables and ropes; and thin-shell structures such as paper and textiles. Each category poses distinct modelling challenges. As Arriola-Rios et al. observe, deformable object manipulation is “still in its infancy” relative to rigid body manipulation [1].
Textiles and cloth represent a particularly demanding case within the simulation of deformable objects. Unlike volumetric soft bodies, they exhibit an extreme mechanical dichotomy as they behave as elastic solids resisting deformation in the plane of the fabric, yet bend almost freely out of plane [4]. This combination makes them prone to buckling under even minimal compression and results in strongly anisotropic behavior due to their woven internal structure [4]. Additionally, most textiles are light enough that aerodynamic forces play a non-negligible role even at the modest motion speeds typical of domestic or industrial manipulation [4], [6]. These combined properties place textiles at the frontier of what current simulation methods can reliably handle, making them a pertinent reference case for the broader field of deformable object modelling.
Simulation approaches
The simulation of deformable objects draws from two historically separate communities: computer graphics, which prioritized visual plausibility and real-time performance, and computational mechanics, which prioritized physical accuracy. Robotic manipulation sits at the intersection of these requirements – accuracy matters for predicting object state and planning manipulation strategies, but computational cost must remain controllable for use in control loops and learning pipelines.
General purpose physics engines such as Gazebo, MuJoCo, PyBullet, SOFA, Webots, V-REP/CoppeliaSim and NVIDIA Isaac Sim have become standard tools for training and evaluating robot controllers [2]. While these platforms excel at rigid-body dynamics, they differ considerably in their support for deformable objects. Some, such as MuJoCo, rely on mesh-based mass-spring models; others, including SOFA, Isaac Sim and PyBullet, additionally support continuum FEM approaches; and NVIDIA FLEX with its robotics-oriented wrapper SoftGym use Position-Based Dynamics – a constraint-based method that prioritises stability and simulation speed over physical accuracy [7]. An emerging alternative, the Material Point Method (MPM), combines particle and grid representations and shows promise for the large-scale deformations encountered in cloth manipulation [7].
Three broad families of physics-based dynamics models underpin these platforms:
- Mesh-based models such as mass-spring systems are intuitive and computationally efficient, but their parameters do not correspond to measurable material constants, behavior is mesh-topology dependent and convergence as the mesh is refined is not guaranteed [3].
- Continuum FEM models are grounded in the equations of continuum mechanics and give model parameters direct physical meaning – stiffness, density and damping can in principle be measured from the real material. They are the most physically accurate option but also the most computationally demanding, and achieving practical simulation speeds requires sophisticated numerical solvers [3].
- Position-Based Dynamics methods operate on particle positions rather than forces, giving unconditional stability suitable for real-time applications, but offer limited physical accuracy and do not allow straightforward calibration to real material properties [7].
Beyond these physics-based approaches, data-driven models have emerged as an alternative paradigm [7]. These learn cloth dynamics directly from data (using graph neural networks, image-based representations or point clouds) and can be conditioned on material properties through domain randomization or explicit property encoding. They offer lower inference cost and greater flexibility than constitutive models, but generalization to unseen material variations remains limited.
Differentiable simulation is a further development within physics-based modelling: by enabling gradients to flow through the physics model, it allows simulator parameters and control policies to be optimized directly against real-world observations, making it particularly valuable for system identification and sim-to-real transfer [7].
Core Challenges
1. Physical Complexity and Material Diversity
Deformable objects exhibit a vast range of mechanical behaviors. Cloth is thin, flexible and nearly inextensible in the plane of the fabric; ropes are one dimensional and subject to large bending; soft tissues are volumetric, viscoelastic and nearly incompressible; food items may undergo plastic deformation or fracture. Each material class may require a tailored modeling approach and no single method handles all cases with equal fidelity.
Textiles in particular expose the limits of current simulation methods. Cloth is highly anisotropic [3] – it resists stretching far more strongly than bending or shearing – and capturing this accurately while maintaining practical simulation speeds is non-trivial. The same piece of fabric can also behave very differently depending on its weave, fiber composition and moisture content, factors that current simulators can only approximate. Most simulators further simplify by modeling cloth as a 2D surface with no thickness [7], an assumption that holds for many tasks but breaks down when simulating folding, layering or stacking. Friction is also typically modeled as isotropic (equal in all directions) whereas real textiles exhibit directional friction behavior due to their woven construction [7], introducing a systematic mismatch between simulation and reality.
2. Contact and Collision Handling
Contact detection and resolution is one of the hardest numerical problems in deformable simulation. For cloth, this involves not only contact between the fabric and external objects (obstacles, robot end-effectors, surfaces) but also self-contact when the fabric folds onto itself. Self-collision introduces topological complexity that, when handled poorly, leads to interpenetration artifacts that are difficult to recover from.
Three broad approaches exist for resolving collisions in deformable simulation. Penalty based methods apply restoring forces when objects overlap, and are simple to implement but require stiff forces to prevent visible interpenetration, which can destabilize the simulation. Impulse based methods correct velocities at the moment of contact and are computationally cheaper, but can struggle with sustained or simultaneous contacts. Constraint based methods enforce non-penetration as a hard constraint alongside the other equations of motion, providing the most physically principled treatment but at greater computational cost [5].
Friction at contact surfaces compounds these difficulties. For textiles, friction must be resolved consistently with the inextensibility constraints of the fabric and the large number of simultaneous contact points, particularly during self-contact, make this a significant numerical challenge.
3. The Sim-to-Real Gap
One of the central difficulties in using simulation for robotic manipulation is the sim-to-real gap – the discrepancy between simulated and real-world behavior that causes controllers trained or designed in simulation to perform poorly when deployed on physical hardware [2]. For rigid bodies this gap is already a known challenge and for deformable objects it is considerably wider. Small errors in material parameters, contact modeling or ignored physical effects can accumulate rapidly into large shape prediction errors, particularly during dynamic manipulation.
A significant contributor to this gap is the cost of accuracy. High-fidelity deformable simulation is expensive and the approximations researchers make to keep it viable, like coarser meshes, simplified contact models or neglected aerodynamics, are often the same approximations that introduce model error. This creates a fundamental tension: the more accurate the simulator the harder it is to run at the speeds required for control or large-scale learning.
Several complementary strategies are being pursued to narrow the gap. One research direction aims to improve the physical fidelity of the simulator itself, reducing the gap at its source by incorporating physical effects that are commonly neglected — for textiles, including aerodynamic forces during dynamic manipulation has been shown to substantially improve trajectory accuracy at manageable computational cost [6].
A second direction uses system identification to fit simulator parameters to real-world observations – either through gradient-based optimization enabled by differentiable simulators, or through black-box methods such as Bayesian optimization and CMA-ES [7].
A third direction, domain randomization, sidesteps precise parameter identification by training controllers across a wide distribution of simulated material properties, relying on the breadth of training to produce policies that transfer robustly [7].
A broader obstacle across all these strategies is the absence of standardized benchmarks: the field uses a heterogeneous collection of tasks and metrics, making it difficult to compare simulators or track progress systematically. Longhini et al. [7] make a step toward taxonomy and evaluation criteria for cloth manipulation, but no widely adopted benchmark suite exists.
4. Parameter Identification and Calibration
Even the most physically accurate simulator is only as useful as its input parameters. Identifying the constitutive parameters of a deformable object (stiffness, damping, density, friction coefficients) from physical measurements is itself a research challenge. Many parameters are difficult to measure directly and must instead be inferred from observed behavior, which can lead to ambiguities. Different parameter combinations may produce similar looking simulations under one set of conditions but diverge under others.
For textiles, this problem is compounded by the number of independent parameters involved. Cloth is anisotropic, meaning separate stiffness and damping values may be needed for the warp, weft and bias directions, and these interact with the contact and friction model in ways that make independent identification difficult. The inextensibility assumption adds further constraints – a model that enforces near-zero in-plane strain removes some degrees of freedom but makes the remaining parameters more sensitive [4].
Several identification strategies are in use. Classical approaches fit simulator parameters by minimizing the discrepancy between simulated and observed trajectories, using gradient-based optimization where differentiable simulators permit it, or black-box methods such as Bayesian optimization and CMA-ES where they do not [7]. Learning based approaches train neural networks to predict material parameters directly from visual observations, offering speed at the cost of generalization [7]. Hybrid methods combine learned perception with physics-based simulation, using the neural network for parameter estimation and the simulator for forward prediction.
Simulator Comparison
The following table summarizes the deformable object capabilities of the simulators most commonly referenced in the robotics manipulation literature. The classification of deformation models follows Longhini et al. [7]; support for deformable objects is drawn from Collins et al. [2].
| Simulator | Deformation model | Deformable object support | Deformable fidelity | Ease of use |
|---|---|---|---|---|
| MuJoCo | Mesh-based (mass-spring) | Yes | ** | ***** |
| SOFA | Mesh-based (mass-spring), FEM | Yes | ***** | * |
| PyBullet | FEM | Yes | ** | **** |
| NVIDIA FLEX / SoftGym | PBD | Yes | *** | *** |
| NVIDIA Isaac Sim | FEM | Yes | **** | *** |
| Gazebo | Rigid body primarily | No | / | *** |
MuJoCo is commonly employed within robotics research, with notable contact stability as a reason for its popularity [2]. For deformable objects it implements mesh-based mass-spring-damper models, representing cloth as a network of interconnected masses and springs [7]. While straightforward to implement, these models’ parameters do not directly correspond to physical material properties and are best suited to small deformations rather than complex elastic effects with high fidelity [7]. MuJoCo does not support full FEM-based cloth simulation, making it more appropriate for cable and tendon modelling than for demanding cloth manipulation tasks.
SOFA (Simulation Open Framework Architecture) is a popular open-source simulator supporting both mesh-based and FEM-based deformable object simulation [2], [7]. It has been applied to FEM simulation of soft robots and provides a ROS bridge and a Soft Robot plugin for modelling and actuation, as well as a plugin for real-time deformable object simulation and haptic control, making it a common choice for medical and surgical robotics research [2]. However, its FEM-based approach is computationally intensive and offers limited GPU acceleration, making it less suited to high-throughput learning training than lighter-weight simulators.
PyBullet supports deformable object and cloth simulation through the Finite Element Method, using continuous domain modelling for deformable objects [2], [7]. It is specifically targeted at robotics research, providing support for reinforcement learning environments, virtual reality integration and forward and inverse kinematics alongside its deformable object capabilities [2]. Its deformable modeling fidelity is nevertheless limited compared to dedicated FEM simulators such as SOFA, and contact handling may become unstable under complex contact conditions, making it better suited to prototyping and simple cloth experiments than to high fidelity cloth manipulation.
NVIDIA FLEX and SoftGym use Position-Based Dynamics, a particle-based approach that satisfies constraints on stiffness, elasticity and collision by directly adjusting particle positions. PBD offers rapid simulation speeds and enhanced stability and can simulate inextensible textiles and shear deformation, though interpreting simulation parameters in terms of physical material properties remains challenging [7]. FLEX is GPU-accelerated and supports distributed GPU simulations [2], while SoftGym is a cloth manipulation benchmark built on FLEX that has been widely used to evaluate learning-based methods for deformable object manipulation [7]. Beyond cloth, FLEX’s unified particle-based solver also handles fluid simulation, enabling cloth-fluid interaction scenarios.
NVIDIA Isaac Sim supports deformable objects through continuous domain (FEM) modelling [2], [7], built on the PhysX 5 engine which incorporates FEM-based soft-body dynamics as well as Position-Based Dynamics for fluids and inflatables. Its GPU-accelerated parallel simulation and built-in domain randomization tools make it particularly well suited for large-scale learning training, though its infrastructure requirements are substantial.
Gazebo lacks native support for deformable objects [2]. While widely used for rigid-body robotics through its ROS integration, it is not well suited to cloth or soft body manipulation research without substantial custom extension.
Conclusion
The simulation of deformable objects for robotic manipulation has advanced considerably, driven by progress in FEM solvers, PBD methods and, more recently, differentiable and data-driven approaches. General purpose robotics simulators now offer basic deformable support, but for demanding material classes (such as textiles) the gap between what simulators can reliably produce and the complexity of real-world behavior remains large. The comparison above reflects this: no single simulator dominates across all criteria – FEM-based tools such as SOFA offer the highest physical fidelity but at significant computational cost and NVIDIA Isaac Sim combines FEM-based deformable support with GPU acceleration but requires substantial infrastructure. PBD-based platforms such as FLEX trade accuracy for speed and scalability, and general purpose simulators such as MuJoCo and PyBullet provide accessible entry points but with limited cloth modeling capability.
Textiles serve as a useful stress test for the field – their combination of near inextensibility, anisotropy, aerodynamic sensitivity and propensity for self-contact exposes weaknesses in every layer of the simulation stack, from constitutive models and contact resolution to parameter identification and sim-to-real transfer. Progress on textile simulation therefore tends to surface challenges relevant to the broader deformable object problem.
The most productive near-term directions appear to be tighter integration of differentiable physics with real-world system identification, more principled handling of self-contact in constraint-based frameworks, and the development of shared benchmarks that allow rigorous comparison across simulators and methods.
Resources
| Tool | Link |
|---|---|
| NVIDIA Isaac Sim | https://developer.nvidia.com/isaac-sim |
| SOFA Framework | https://www.sofa-framework.org/ |
| PyBullet | https://github.com/bulletphysics/bullet3 |
| MuJoCo | https://mujoco.org/ |
| NVIDIA Flex | https://github.com/NVIDIAGameWorks/FleX |
| Genesis | https://github.com/Genesis-Embodied-AI/Genesis |
| Newton Dynamics | https://github.com/newton-physics/newton |
| NVIDIA Warp | https://github.com/NVIDIA/warp |
| DiffTaichi | https://github.com/taichi-dev/difftaichi |
| SoftGym | https://github.com/Xingyu-Lin/softgym |
| ClothGym | https://github.com/columbia-ai-robotics/clothgym |
| Isaac Lab | https://github.com/NVIDIA-Omniverse/IsaacLab |
References
[1] V. E. Arriola-Rios, P. Guler, F. Ficuciello, D. Kragic, B. Siciliano, and J. L. Wyatt, “Modeling of Deformable Objects for Robotic Manipulation: A Tutorial and Review,” Front. Robot. AI, vol. 7, Sep. 2020, doi: 10.3389/frobt.2020.00082.
[2] J. Collins, S. Chand, A. Vanderkop, and D. Howard, “A Review of Physics Simulators for Robotic Applications,” IEEE Access, vol. 9, pp. 51416–51431, 2021, doi: 10.1109/ACCESS.2021.3068769.
[3] A. Nealen, M. Müller, R. Keiser, E. Boxerman, and M. Carlson, “Physically Based Deformable Models in Computer Graphics,” Computer Graphics Forum, vol. 25, no. 4, pp. 809–836, Dec. 2006, doi: 10.1111/j.1467-8659.2006.01000.x.
[4] F. Coltraro, J. Amorós, M. Alberich-Carramiñana, and C. Torras, “An inextensible model for the robotic manipulation of textiles,” Applied Mathematical Modelling, vol. 101, pp. 832–858, Jan. 2022, doi: 10.1016/j.apm.2021.09.013.
[5] F. Coltraro, J. Amorós, M. Alberich-Carramiñana, and C. Torras, “A novel collision model for inextensible textiles and its experimental validation,” Applied Mathematical Modelling, vol. 128, pp. 287–308, Apr. 2024, doi: 10.1016/j.apm.2024.01.030.
[6] F. Coltraro, J. Amorós, C. Torras, and M. Alberich-Carramiñana, “A practical aerodynamic model for dynamic textile manipulation in robotics,” Mechanism and Machine Theory, vol. 209, p. 105993, Jul. 2025, doi: 10.1016/j.mechmachtheory.2025.105993.
[7] A. Longhini et al., “Unfolding the Literature: A Review of Robotic Cloth Manipulation,” Jul. 16, 2024, arXiv: arXiv:2407.01361. doi: 10.48550/arXiv.2407.01361.