As-Rigid-As-Possible (ARAP) Energy

• ARAP energy is a metric that quantifies deviation from local rigidity by assigning an optimal rotation per mesh element, typically solved via SVD.
• It underpins techniques in dynamic 3D reconstruction, non-rigid shape editing, and neural field deformations by ensuring locally consistent transformations.
• Extensions such as smooth and rotation-invariant variants improve continuity, enhance real-time interactivity, and support seamless integration into deep learning models.

The As-Rigid-As-Possible (ARAP) energy is a foundational concept for regularizing non-rigid deformations in computational geometry, computer vision, and shape processing. ARAP energy quantifies departure from local rigidity during deformation, enforcing deformations that minimize deviations from ideal rigid-body motion at a local or patchwise scale. Originating in surface and volumetric mesh deformation, ARAP regularization now underpins a broad family of algorithms for editing and reconstructing 3D data, deformable shape modeling, tracking, neural field deformation, and deep shape generation.

1. Core Formulation of ARAP Energy

At its core, the ARAP energy is defined on a discretized domain (usually a polygonal mesh or a tetrahedralization) and assigns to each mesh element or node a locally optimal rigid transformation. The classical ARAP energy for a domain $\Omega$ and a deformation $f:\Omega\to\mathbb R^3$ is, in the continuous setting,

$$E_{\rm cont}[f] = \int_{\Omega}\min_{R(x)\in \mathrm{SO}(3)} \|\nabla f(x)-R(x)\|_{F}^2~dx,$$

where $R(x)$ is the best-fitting local rotation and $\|\cdot\|_F$ denotes the Frobenius norm.

In the standard discrete case, for a mesh $(V,E)$ with rest positions $g=(g_1,\dots,g_n)\in\mathbb R^{3n}$ and deformed positions $x=(x_1,\dots,x_n)\in\mathbb R^{3n}$, together with local rotations $O_i\in \mathrm{SO}(3)$ per vertex, the discrete ARAP energy reads

$$f_R(g, x) = \min_{\{O_i\in \mathrm{SO}(3)\}} \sum_{(i,j)\in E} \|(O_i-I_3)(g_i-g_j)-(x_i-x_j)\|^2$$

or, with edge weights $w_{ij}$,

$$E_{\rm ARAP} = \sum_{(i,j)\in E} w_{ij} \min_{O_i} \|(O_i-I_3)(g_i-g_j)-(x_i-x_j)\|^2$$

(Kaji, 2016, Huang et al., 2021).

The ARAP energy thus penalizes local deviation from rigidity by comparing the deformed edge vectors to the original edges transformed by local optimal rotation. The minimization over $O_i$ can be solved in closed form per-vertex via SVD (Procrustes alignment).

2. Algorithmic Schemes and Optimization

Minimizing ARAP energy is typically performed using a local-global optimization framework:

• Local step: Given deformed positions, solve for optimal $O_i$ per node/element, usually via Procrustes (SVD).
• Global step: With rotations fixed, solve for new positions $x_i$ by minimizing the quadratic energy, leading to a sparse linear system.

This framework elegantly generalizes to both surface and volumetric meshes through appropriate definitions of neighboring sets, weighting (e.g., cotangent weights for surface meshes, volume weights for tetrahedra), and discretization schemes (Kaji, 2016, Oehri et al., 17 Jan 2025).

Alternative ARAP energies include translation-invariant and rotation-invariant forms, the latter focusing solely on non-rigid (shear) deformations: for example, $E_S$ compares only the symmetric factors in polar decompositions, achieving full rotation invariance (Kaji, 2016).

Recent methods extend the classical alternating minimization with higher-order smoothness (Oehri et al., 17 Jan 2025) or embed ARAP as a differentiable objective in neural optimization pipelines (Baieri et al., 2024, Božič et al., 2020).

3. ARAP in Dynamic Scene Reconstruction and Regularization

ARAP energy is critical in regularizing piecewise-rigid dynamic 3D reconstruction, as in the "Superpixel Soup" model (Kumar et al., 2019). Here, an input frame is over-segmented into $N$ superpixels, each modeled as a small planar patch with rigid motion. The ARAP term is defined over a $K$-nearest neighbor (KNN) graph of superpixel anchors and consists of three terms:

1. Smoothness of rotation: $\|R_i-R_k\|_F$ between neighboring patches,
2. Smoothness of translation magnitude: $\|\lambda_i\hat t_i - \lambda_k\hat t_k\|_2$,
3. Preservation of inter-anchor distances: $$|\|\tilde x_{a_i}-\tilde x_{a_k}\|_2 - \|\tilde x'_{a_i}-\tilde x'_{a_k}\|_2|$$.

The ARAP regularizer jointly penalizes non-smooth local motions and enforces global as-rigid-as-possible behavior across the scene, critically resolving scale ambiguity and preserving geometric fidelity under large non-rigid motion (Kumar et al., 2019).

In ablation, using only $E_{\rm arap}$ achieves mean relative depth error (MRE) ≈ 0.2248, and omitting ARAP leaves MRE above 0.27, demonstrating that ARAP is essential for aligning local scales and maintaining physical plausibility.

4. Extensions: Smoothness, Higher-Order, and Neural Domain Generalizations

A significant issue in classical ARAP is the introduction of discontinuities ("spikes") near positional constraints. The "smooth ARAP" (or higher-order ARAP) remedies this by adding a second-order term penalizing non-rigid changes in the (cotan) Laplacian of vertex positions (Oehri et al., 17 Jan 2025). The cumulative energy blends $E_{\rm ARAP}$ and a Laplacian-aligned smoothness energy: $$E(V',\{R_v\}) = (1-\lambda)\,E_{\rm ARAP}(V') + \lambda\,E_{\rm smooth}(V'),$$ where $$E_{\rm smooth}(V') = \sum_{v\in V} A_v \|\ell_v' - R_v\ell_v\|^2$$ for Laplacian vectors $\ell_v$. For $\lambda\approx 0.95$, this eliminates spikes and ensures $C^1$ continuity at handles, enabling robust, interactive editing even with single-point constraints (Oehri et al., 17 Jan 2025).

In the neural domain, ARAP-inspired losses are crucial for learning plausible deformable generators. The ARAPReg approach constructs a spectral penalty on the Hessian of the ARAP energy projected into the latent tangent space, penalizing only non-rigid (shape, not pose) variations while enabling closed-form, differentiable integration into VAE or auto-decoder frameworks (Huang et al., 2021).

Likewise, the "Implicit-ARAP" framework generalizes ARAP energy to neural signed distance fields (SDFs) by building local patch meshes sampled on isosurfaces (e.g., level sets) and computes per-patch ARAP error as a fully differentiable objective. This permits ARAP-constrained deformation in high-resolution neural fields with unprecedented efficiency and topology independence (Baieri et al., 2024).

5. Robustness, Invariance, and Tetrisation

Classical ARAP is only translation-invariant; to fully decouple from global rotations (which is critical for shape blending, morphing, or when input poses differ), a rotation-invariant variant penalizes only deviation in the symmetric part (shear/stretch) of the per-element Jacobians after polar decomposition (Kaji, 2016). The rotation-invariant ARAP energy is

$$E_S(V',\{C_i\}) = \sum_{i=1}^n \|S(\hat A_i(V')) - S(C_i)\|_F^2$$

where $S(X)$ is the symmetric factor in the polar decomposition $X=RS$. This is crucial for blending shapes with large misorientations without artifacts and for applications such as interpolating between keyframes in different poses.

Applying ARAP energies to surface meshes requires embedding them in a volumetric (tetrahedral) structure, known as tetrisation. Variants such as face-normal, edge-normal, or vertex-normal tetrisations encode local geometry and adjacency. The choice of tetrisation affects not just the Laplacian structure but also deformation quality—edge- or vertex-normal pockets capture curvature and yield better preservation of sharp features (Kaji, 2016).

6. ARAP in Deep Learning, Fields, and High-Resolution Applications

Tractable and efficient ARAP energy approximations extend its utility to neural settings. In "Implicit-ARAP," deformation is parameterized as a neural field $d_\phi$ with per-point predicted rotation $R_\phi(x)$ and translation $t_\phi(x)$. The energy over a patch mesh $k$ is

$$L_{\rm arap}^k = \sum_{(v_i,v_j)\in E_k} w_{ij} \|[d_\phi(v_i)-d_\phi(v_j)] - [R_\phi(v_i)(v_i-v_j)]\|^2,$$

aggregated over many small patches (Baieri et al., 2024). Unlike traditional methods, there are no global sparse linear solves or SVDs on the full mesh, and cost per step is independent of mesh resolution.

This approach allows as-rigid-as-possible constraints to be imposed directly on implicit representations, yielding efficient, robust, and scalable deformations applicable even on neural SDFs—enabling shape generation, deformation, and editing at a scale and generality not feasible with explicit mesh-based ARAP.

7. Impact, Practical Considerations, and Quantitative Results

ARAP energy is now foundational in geometric modeling, non-rigid structure-from-motion, shape registration, and neural shape generation. Accurate, robust scale and shape recovery, avoidance of nonphysical distortions, and real-time interactivity on high-resolution models are its principal benefits.

Performance evaluations show:

• "Smooth ARAP" eliminates spikes at handles and achieves faster convergence (e.g., 25–200 iterations versus 200–400 for classical ARAP), and real-time solve times (e.g., 10–30 Hz at 200k verts) (Oehri et al., 17 Jan 2025).
• "Implicit-ARAP" achieves 133$\times$ speed and order-of-magnitude memory gains versus dense-grid methods, handles meshes of up to $10^6$ faces, and completely bypasses dependence on mesh quality, enabling deformations even from SDFs without explicit meshing (Baieri et al., 2024).
• In the "Superpixel Soup" model for dynamic scene reconstruction, the ARAP term alone reduces mean relative depth-error from >0.27 to ≈0.22, and combined with smoothness/orientation/continuity terms, MRE drops to ≈0.16 (Kumar et al., 2019).

A plausible implication is that as-rigid-as-possible energy regularization will remain central in the integration of classical geometry processing and modern neural field methods, particularly as topologically flexible, resolution-independent, and fully differentiable formulations become standard.

---

Key References:

• (Kumar et al., 2019) Superpixel Soup: Monocular Dense 3D Reconstruction of a Complex Dynamic Scene
• (Baieri et al., 2024) Implicit-ARAP: Efficient Handle-Guided Deformation of High-Resolution Meshes and Neural Fields via Local Patch Meshing
• (Oehri et al., 17 Jan 2025) Higher Order Continuity for Smooth As-Rigid-As-Possible Shape Modeling
• (Kaji, 2016) Tetrisation of triangular meshes and its application in shape blending
• (Huang et al., 2021) ARAPReg: An As-Rigid-As Possible Regularization Loss for Learning Deformable Shape Generators
• (Božič et al., 2020) Neural Non-Rigid Tracking