EquiForm: Robust Equivariance & Equilibrium

• EquiForm is a collection of frameworks that leverage symmetry principles, including equivariance and equilibrium, across geometric deep learning, finite element analysis, and differential geometry.
• It features an SE(3)-equivariant policy learning pipeline with geometric denoising and contrastive alignment to achieve noise-robust robotic manipulation from 3D point clouds.
• The approach significantly improves performance in simulation and real-world tasks, while ensuring accurate stress approximations and scale invariance in advanced computational models.

EquiForm refers to a set of mathematical and computational frameworks that leverage equivariance principles, scale invariance, or equilibrium constraints to achieve robust geometric or physical reasoning in diverse domains. The name has been adopted for advanced methods in geometric deep learning, equilibrium finite element formulations, and scale-invariant geometry, each with distinct formulations and applications. This entry surveys the three primary meanings of EquiForm, with particular emphasis on SE(3)-equivariant policy learning for 3D point cloud-based manipulation, as well as related developments in structural mechanics and differential geometry.

1. SE(3)-Equivariant Policy Learning with EquiForm

EquiForm is a noise-robust SE(3)-equivariant policy learning framework for robotic manipulation from 3D point clouds (Zhang et al., 24 Jan 2026). The approach is founded on the requirement that observation-to-action mappings in robot policy learning are equivariant with respect to rigid body transformations in SE(3) (the group of 3D rotations and translations). For an observation $O = (X, \mathbf s)$, where $X \in\mathbb R^{3\times N}$ is a point cloud and $\mathbf s$ is a proprioceptive state, the policy $\pi : O \mapsto \mathbf a$ is SE(3)-equivariant if

$\pi(T \triangleright O) = T \triangleright \pi(O)$

for all $T \in \mathrm{SE}(3)$, where $T \triangleright X = RX + \mathbf b\mathbf 1^\top$ and $$T \triangleright \mathbf a = R\mathbf a + \mathbf b$$.

SE(3)-equivariant encoders ($\Psi$) are used to canonicalize point cloud observations: $\hat X = \Psi(X)^{-1} \triangleright X.$ This allows the policy to output actions in a canonical pose and map them back to the world frame, preserving the desired equivariance.

2. Handling Noise-Induced Equivariance Deviations

In realistic settings, point cloud observations are distorted by noise, occlusions, and pose perturbations, breaking SE(3) equivariance. The stochastic noise operator $\delta$ applied to a point cloud ($X_\delta = \delta(X)$) is not SE(3)-equivariant: $\delta(TX) \neq T\delta(X),$ and thus

$$\Psi\bigl(\delta(TX)\bigr) \neq T\Psi\bigl(\delta(X)\bigr).$$

Consequently, noisy versions of otherwise equivalent observations can yield inconsistent latent representations and policy outputs. This "equivariance deviation" is a primary source of generalization failure in existing equivariant neural architectures for robotic manipulation (Zhang et al., 24 Jan 2026).

3. Geometric Denoising in EquiForm

EquiForm introduces an explicit geometric denoising module to address the above challenge. Each noisy point $\mathbf x_\delta$ is projected onto a locally estimated surface normal and smoothed in the tangent direction:

• Normal-direction projection: Computes the local neighborhood mean $\bar{\mathbf x}_\delta$, estimates the normal $\mathbf n_\delta$ (via PCA), then projects using

$$\mathbf x = \mathbf x_\delta - \mathbf n_\delta\mathbf n_\delta^\top(\mathbf x_\delta - \bar{\mathbf x}_\delta).$$

• Tangent-direction smoothing: Updates each $\mathbf x$ using

$$\mathbf x^\star = \mathbf x - (I - \mathbf n\mathbf n^\top)(\mathbf x - \bar{\mathbf x}).$$

The denoised point cloud $X^\star$ serves as the foundation for robust SE(3)-equivariant encoding and policy prediction.

4. Contrastive Equivariant Alignment Objective

Residual inconsistencies due to partial views and occlusion are addressed by a contrastive InfoNCE loss on the encoder's latent features (Zhang et al., 24 Jan 2026). For each denoised cloud $X^\star$ and its augmented view $\tilde X^\star$, the InfoNCE loss is defined as: $$\mathcal L_{\rm NCE} = -\sum_i \log \frac{\exp(\mathrm{sim}(\mathbf z_i, \tilde{\mathbf z}_i)/\tau)}{\sum_j\exp(\mathrm{sim}(\mathbf z_i, \tilde{\mathbf z}_j)/\tau)},$$ where $\mathbf z = \Psi(X^\star)$, $\tilde{\mathbf z} = \Psi(\tilde X^\star)$, and similarity is measured by cosine similarity. Minimizing $\mathcal L_{\rm NCE}$ regularizes representations to remain consistent under rigid transformations and noise, preserving the SE(3) structure.

5. Policy Learning Pipeline and Experimental Results

The EquiForm pipeline at each timestep processes a raw cloud, applies geometric denoising, extracts features via an SE(3)-equivariant encoder, and predicts a canonical action which is mapped back to the original pose. The total training loss combines standard behavior cloning and the contrastive equivariant alignment term. The action head is implemented as a diffusion model.

Empirical evaluations cover 16 simulation tasks and 4 real-robot manipulation tasks (Zhang et al., 24 Jan 2026). Results are summarized below.

Task domain	DP3 (non-equivariant)	Canonical Policy	EquiForm
Simulation (avg)	42.3%	56.6%	66.6%
Real Robot (avg)	12.5%	17.5%	45%

EquiForm improves average success in simulation by 17.2 percentage points compared to Canonical Policy and by 28.1 points in real-world benchmarks. Ablation shows the geometric denoising and contrastive modules yield complementary gains.

6. EquiForm in Finite Element Equilibrium

EquiForm also designates a high-order mixed spectral element method for elastic equilibrium (Olesen et al., 2016). This formulation uses integrated traction components as degrees of freedom for the stress field, directly enforcing equilibrium of forces pointwise (if body forces are polynomial), and enforcing equilibrium of moments weakly via Lagrange multipliers. Spectral element bases are constructed to guarantee interelement traction continuity and pointwise satisfaction of equilibrium constraints. Practical outcomes include monotonic convergence of the complementary energy functional, physically accurate stress and displacement approximations, and resilience to mesh singularities on both orthogonal and curvilinear grids.

7. EquiForm in Differential Geometry

In the context of 4D Minkowski space, EquiForm refers to the equiform differential geometry of curves (Abdel-Aziz et al., 2015). For a smooth spacelike curve $\alpha: I \to E_4$, the equiform parameter $\sigma = \int \kappa_1(s)\,ds$ and the associated equiform frame yield equiform curvatures: $$\tilde{k}_1 = -\frac{\kappa_1'}{\kappa_1^3},\quad \tilde{k}_2 = \frac{\kappa_2}{\kappa_1},\quad \tilde{k}_3 = \frac{\kappa_3}{\kappa_1}.$$ Plane curves are characterized by $\tilde{k}_2 \equiv 0$, hyperplane curves by $\tilde{k}_3 \equiv 0$, and general helices (constant slope) by both $\tilde{k}_2, \tilde{k}_3$ constant. These equiform invariants are relevant for relativistic kinematics, computer vision (scale-invariant curve matching), and the study of self-similar motions.

---

The unifying principle across these interpretations of EquiForm is the explicit modeling or enforcement of symmetry-related structure—be it SE(3)-equivariance in policy learning, equilibrium in numerical PDE solvers, or scale invariance in the differential geometry of curves. Each instantiation leverages this structure for heightened robustness, generalization, and physical consistency in geometric computation and learning (Zhang et al., 24 Jan 2026, Olesen et al., 2016, Abdel-Aziz et al., 2015).