Position-Based Dynamics (PBD)

• Position-Based Dynamics (PBD) is a simulation method that directly manipulates particle positions using iterative constraint projections rather than force-based equations.
• It achieves unconditional stability and efficiency via a Gauss–Seidel projection loop, making it ideal for real-time graphics, robotics, and interactive environments.
• Extensions like XPBD and differentiable frameworks further enhance PBD’s capabilities, enabling accurate modeling of compliant constraints, rigid bodies, and data-driven materials.

Position-Based Dynamics (PBD) is an explicit, constraint-projection approach for simulating the dynamics of particles, deformable bodies, and rigid bodies via direct manipulation of positions, as opposed to integrating force-based equations of motion. PBD has achieved prominence in real-time computer graphics, robotics, and interactive environments due to its unconditional stability, simplicity, and composable constraint modeling. The core idea is that mechanical behavior is encoded through geometric constraint functions imposed on the particle positions, and time stepping alternates prediction under external forces with iterative projection onto the constraint manifold. Recent developments generalize PBD to compliant constraints (XPBD), rigid-body systems, differentiation, continuum inelasticity, and data-driven materials, giving PBD a unique position among modern simulation toolkits.

1. Core Algorithmic Structure of PBD

PBD models a physical system as a set of $N$ point-mass particles with positions $\{p_i \in \mathbb{R}^3\}$ and masses $m_i$, or more generally, as bodies with positions and orientations. Mechanical behavior is encoded by $M$ geometric constraints $C_k(\mathbf{p}) = 0$ or $C_k(\mathbf{p}) \geq 0$, applied to local subsets of points or bodies.

The time-integration cycle for each frame consists of:

• Prediction of unconstrained positions via explicit or symplectic Euler:

$$p_i^* = p_i^t + \Delta t\,v_i^t + \Delta t^2\,m_i^{-1} f_i^t$$

where $f_i^t$ includes external forces.

• Constraint projection: Iteratively project $\{p_i^*\}$ to satisfy all constraints using (block-)Gauss–Seidel:

$\Delta p_{i} = -w_i\,\lambda\,\nabla_{p_i}C$

with

$$\lambda = \frac{C(p^*)}{\sum_j w_j\,\|\nabla_{p_j}C\|^2 + \varepsilon }$$

Here $w_i=1/m_i$ and $\varepsilon$ guards against numerical singularity.

• Velocity update:

$v_i^{t+1} = \frac{p_i^{t+1} - p_i^t}{\Delta t}$

Common constraints include distance (edge) constraints, volume (tetrahedron) preservation, bending, collision/contact, and per-cluster shape matching (Liu et al., 2020). Efficient real-time performance is achieved via a fixed number of constraint solver iterations per timestep, typically on the order of 5–10 (Liu et al., 2020, Seabra et al., 2023).

2. Mathematical Foundations: Projective Constraint Solvers

At the constraint level, PBD is mathematically a nonlinear block Gauss–Seidel solver for a projection onto the feasible set defined by all $C_k$. Linearizing $C_k$ at the current positions, one applies a minimal-norm (usually mass-weighted) correction $\Delta p$ to achieve $C_k \approx 0$:

$$\Delta p_i = -\frac{w_i\,C_k}{\sum_j w_j \|\nabla_{p_j}C_k\|^2}\nabla_{p_i}C_k$$

This is equivalent to a single step of an affine-IQP subproblem per constraint, then immediate update of positions in-place (Gauss–Seidel rather than Jacobi).

XPBD (extended/compliant PBD) further augments each constraint with a compliance parameter $\alpha$:

$$\Delta \lambda = -\frac{C_k + \tilde{\alpha}\lambda}{J_k + \tilde{\alpha}}, \quad \tilde{\alpha} = \alpha / \Delta t^2$$

where $J_k = \nabla C_k^T M^{-1} \nabla C_k$ (Stuyck et al., 2023, Seabra et al., 2023). This enables proper time-step-independent control of stiffness and critical for robustly modeling compliant and near-rigid material regimes (Seabra et al., 2023, Yu et al., 19 May 2024).

Recent extensions generalize the projection step to optimal area/orientation for triangles (Duque et al., 2020), per-cluster shape-matching (Caccamo et al., 2018), and arbitrary nonlinear energies for generalized PBD (GPBD) (Chaudhary et al., 28 Nov 2025). In each case, the constraint solve becomes a small-dimension optimal projection or Newton/minimization subproblem.

A significant theoretical result is the rigorous proof of convergence of sequential, prox-regular constraint projections applied in PBD to the solution of the corresponding differential inclusion in the limit of small time step (Plunder et al., 2023).

3. Extensions: XPBD, Rigid Bodies, Nonlinear and Differentiable PBD

XPBD: The XPBD framework formally derives the constraint update as a backward-Euler step on a quadratic constraint potential, introducing compliance $\alpha$ for soft constraints, and retaining Lagrange multiplier history for accurate time-discretization (Stuyck et al., 2023). This removes timestep-dependence from stiffness and enables physically meaningful softness and damping regimes.

Rigid-body PBD: Rigid-body extension (PBRBD) augments particles with rotational DOFs (usually quaternions), angular velocities, inertia tensors, and supports constraints such as joints, hinges, and orientation limits. Corrections to positions at non-center-of-mass points induce both translational and rotational updates, handled via back-projected offsets and quaternion increments (Seabra et al., 2023). Contacts/collisions and joint limits are modeled as positional constraints with compliance, yielding a unified solver for soft and rigid bodies in the same simulation framework.

Generalized (GPBD) and Nonlinear Constraints: GPBD reformulates implicit time integration in terms of individual force terms, enabling Gauss–Seidel updates even for arbitrary nonlinear (e.g., data-driven) hyperelastic materials by projecting onto the minimal-norm displacement in each force/strain space (Chaudhary et al., 28 Nov 2025). This subsumes XPBD and extends the domain of applicability of PBD while retaining unconditional stability.

Differentiable PBD: Analytical gradients for both hard and compliant/XPBD constraints can be derived for gradient-based optimization (Stuyck et al., 2023, Liu et al., 2022, Zhong et al., 2022). The full forward simulation is unrolled, and adjoint or automatic differentiation propagates gradients through constraint projections, collisions, and external force sequences for parameter identification, control, or inference. Contact projections introduce piecewise-smooth mapping; appropriate regularization or XPBD compliance aids in restoring meaningful gradients for learning and control tasks.

4. Application Domains and Real-World Integrations

PBD's modular architecture allows integration in a range of domains:

• Deformable body modeling: Soft tissue simulation in surgical robotics leverages PBD for interactive, vision-driven real-to-sim registration and tracking, achieving accuracy gains in occluded regions and low computational overhead (Liu et al., 2020).
• Crowd and multi-agent simulation: PBD supports short-range and anticipatory long-range collision constraints, friction, and coalescence via Lagrangian position constraints, efficiently simulating hundreds of thousands of interacting agents (Weiss et al., 2018).
• Cloth, fluids, and avatars: Unified point-based representations with constraint libraries for stretching, bending, volume preservation, air-mesh inversion prevention, and collision support enforce realistic physical behavior in virtual avatars, garments, and multi-phase fluids (Sasaki et al., 5 Dec 2024, Feng et al., 27 Jan 2024).
• Articulated and rigid-body robotics: PBD underpins recent differentiable simulators for robot model identification, motion control, optimal design, and impedance control in human-in-the-loop real hardware settings (Liu et al., 2022).
• Rope/rod and volumetric muscle representation: Position-based rods with scale degrees of freedom model large deformation, muscle extrusion, and skinning to surface meshes in anatomically realistic contexts; XPBD enables robust simulation of both inextensible and extensible elastic rods with differentiable solvers (Angles et al., 2019, Liu et al., 2022).
• Continuum inelasticity: XPBD extended with SPH-based kernel estimators and in-the-loop plasticity/enhanced constitutive models simulates elastoplastic, viscoplastic, and granular continua, bridging the gap with MPM-style solvers (Yu et al., 19 May 2024).

5. Algorithmic Performance, Stability, and Practical Considerations

PBD and its extensions are characterized by:

• Unconditional stability: Time step limitations are minimal; the Gauss–Seidel projection loop suppresses numerical blow-up and avoids the requirement for global matrix assembly or inversion.
• Efficiency and scalability: Per-constraint, per-particle updates suit highly data-parallel GPU implementations. Sub-millisecond per-frame solver times are routinely achieved for $O(10^4)$–$O(10^6)$ degrees of freedom (Seabra et al., 2023, Angles et al., 2019, Stuyck et al., 2023).
• Iteration count: For convergent (quasi-static) simulations or high-fidelity physical modeling, more than 10–20 global sweeps may be necessary (Duque et al., 2020, Chen et al., 2023). In real-time graphics, 5–10 iterations suffice for visually plausible behavior. Successive over-relaxation (SOR) and Chebyshev acceleration further improve convergence (Chen et al., 2023).
• Limitation in constraint stacking: For very large or over-constrained graphs (e.g., tall rigid-body piles), Gauss–Seidel-style PBD may suffer from slow convergence or order artifacts. Methods such as constraint compliance, Jacobi coloring, or position-based nonlinear Gauss–Seidel alleviate these issues (Seabra et al., 2023, Chen et al., 2023).
• Hybrid and specialized projectors: For degenerate or highly nonlinear constraints (e.g., area/orientation for nearly collapsed triangles), closed-form or adaptive specialized projectors boost reliability and accuracy (Duque et al., 2020).

6. Theoretical Properties and Research Frontiers

The rigorous mathematical foundation for PBD has recently been established for first-order systems, demonstrating uniform convergence to solutions of differential inclusions with prox-regular constraint manifolds (Plunder et al., 2023).

Research frontiers include:

• Generalization to arbitrary nonlinear forces/energies: GPBD (Chaudhary et al., 28 Nov 2025) and position-based collocation/time integration (Pan et al., 2017) enable the handling of arbitrary, even data-driven, constitutive models at competitive performance versus sparse Newton solvers.
• Integration with machine learning and differentiable programming: Exact adjoint methods for XPBD permit gradient-based material/model parameter estimation, policy learning, and real-world robot control (Stuyck et al., 2023, Liu et al., 2022, Zhong et al., 2022).
• Plasticity, damage, and inelasticity: XPBI introduces in-the-loop implicit return mapping for constitutive models, extending XPBD to elastoplastic and granular materials (Yu et al., 19 May 2024).
• Meshless, hybrid, and multi-physics problems: Extension to meshless shape-matching and coupling with SPH/3D Gaussian Splatting-based rendering yields unified simulation and synthesis frameworks for virtual environments (Caccamo et al., 2018, Feng et al., 27 Jan 2024).

Open challenges persist in the formal analysis of higher-order (second-order, rigid-body impact) PBD, as well as in robust and efficient handling of large, densely coupled constraint graphs at scale.

---

References

• Real-to-Sim Registration of Deformable Soft Tissue with Position-Based Dynamics for Surgical Robot Autonomy (Liu et al., 2020)
• An Optimal Triangle Projector with Prescribed Area and Orientation, Application to Position-Based Dynamics (Duque et al., 2020)
• Position-Based Nonlinear Gauss-Seidel for Quasistatic Hyperelasticity (Chen et al., 2023)
• Survey of Rigid Body Simulation with Extended Position Based Dynamics (Seabra et al., 2023)
• XPBI: Position-Based Dynamics with Smoothing Kernels Handles Continuum Inelasticity (Yu et al., 19 May 2024)
• DiffXPBD : Differentiable Position-Based Simulation of Compliant Constraint Dynamics (Stuyck et al., 2023)
• Towards Generalized Position-Based Dynamics (Chaudhary et al., 28 Nov 2025)
• VIPER: Volume Invariant Position-based Elastic Rods (Angles et al., 2019)
• Parameter Identification and Motion Control for Articulated Rigid Body Robots Using Differentiable Position-based Dynamics (Liu et al., 2022)
• Differentiable Robotic Manipulation of Deformable Rope-like Objects Using Compliant Position-based Dynamics (Liu et al., 2022)
• Position-Based Multi-Agent Dynamics for Real-Time Crowd Simulation (Weiss et al., 2018)
• Gaussian Splashing: Unified Particles for Versatile Motion Synthesis and Rendering (Feng et al., 27 Jan 2024)
• Active Perception and Modeling of Deformable Surfaces using Gaussian Processes and Position-based Dynamics (Caccamo et al., 2018)
• Time Integrating Articulated Body Dynamics Using Position-Based Collocation Method (Pan et al., 2017)
• Convergence proof for first-order position-based dynamics: An efficient scheme for inequality constrained ODEs (Plunder et al., 2023)
• Position based dynamic of a particle system: a configurable algorithm to describe complex behaviour of continuum material starting from swarm robotics (dell'Erba, 2020)
• PBDyG: Position Based Dynamic Gaussians for Motion-Aware Clothed Human Avatars (Sasaki et al., 5 Dec 2024)
• Differentiable Physics Simulations with Contacts: Do They Have Correct Gradients w.r.t. Position, Velocity and Control? (Zhong et al., 2022)