PoseShield: Neural Collision Fields for Human Self-Collision Resolution¶

Conference: ECCV 2026
arXiv: 2606.29686
Code: None
Area: 3D Vision
Keywords: SMPL self-collision resolution, neural collision fields, Eikonal regularization, constrained optimization, human motion generation

TL;DR¶

PoseShield learns a neural collision field with Eikonal regularization (approximating pose-space SDF) within the SMPL pose space, embedding it as a differentiable constraint function in gradient-based constrained optimizers like SLSQP. This achieves post-processing correction from self-penetrating poses to the nearest collision-free poses. This method scales to motion sequence collision resolution without retraining downstream motion generation models, achieving a 95.8% success rate on the self-built HwC dataset, significantly outperforming COAP (44.6%).

Background & Motivation¶

Background: The SMPL parametric human model is the standard geometric representation for human pose estimation and motion generation. However, whether from monocular reconstruction (self-penetrations are widespread in datasets like PROX) or random motion synthesis (outputs from diffusion models like MDM and MoMask), the generated human meshes frequently exhibit self-collision, which undermines physical plausibility. Existing post-processing collision resolution methods fall into two categories: one operates in the mesh space (penalizing energy, interior point methods), which requires a collision-free reference configuration and optimizes vertices rather than pose parameters, making it impossible to directly embed into SMPL-based learning pipelines; the other uses a differentiable penetration loss or a learned collision classifier as a soft constraint in the pose space, but lacks guarantees of gradient regularity, leading to numerical instability in gradient-based constrained optimizers.

Limitations of Prior Work: Pure mesh-based methods cannot adapt to SMPL pipelines where pose parameters are the optimization variables. Meanwhile, existing pose-space methods (such as COAP's occupancy field sampling penalty and N-Penetrate's neural classifier), though differentiable, may suffer from vanishing gradients of the constraint function near the constraint boundaries (violating the LICQ), causing standard constrained optimizers like SLSQP to fail by getting trapped in infeasible stationary points.

Key Challenge: Collision detection is inherently a binary, non-differentiable discrete function, whereas constrained optimizers require constraint functions to be at least \(C^1\) smooth and satisfy the LICQ (non-zero gradients at the boundary). How to maintain collision discrimination ability while ensuring that the gradient norm of the neural constraint function is uniformly bounded throughout the pose space is the key to making constrained optimizers work reliably.

Goal: (1) Learn a differentiable collision constraint function \(g(\theta)\) defined on the SMPL pose space, whose sign accurately indicates the collision state and whose gradient is non-zero globally; (2) use this function as a constraint to solve for the nearest collision-free pose using SLSQP; (3) reuse the same constraint function for motion sequence collision resolution, decoupled from the motion generator.

Key Insight: The authors discover that the LICQ condition (non-zero constraint gradient) in constrained optimization theory naturally maps to the Eikonal equation—a signed distance function (SDF) satisfies a gradient norm of 1 everywhere. Therefore, if a neural collision field can be trained to approximate an SDF (satisfying \(\|\nabla g\| \approx 1\)), it will naturally satisfy the LICQ, thereby providing global convergence guarantees for gradient-based constrained optimizers. This is the core insight of this paper.

Core Idea: Train a pseudo-SDF in the pose space with Eikonal regularization as a differentiable collision constraint, replacing binary collision detection in mesh space, to enable the constrained optimizer (SLSQP) to reliably converge from any self-penetrating pose to the nearest collision-free pose.

Method¶

Overall Architecture¶

The complete pipeline of PoseShield is divided into two phases: training and inference. In the training phase: a large number of poses are sampled within the SMPL 6D rotation pose space (21 joints \(\times\) 6 = 126 dimensions), annotated with binary labels using the FCL exact collision detector, and a 12-layer MLP \(g(\theta)\) is trained with both sign supervision and Eikonal/TD regularization to make its output approximate the signed distance from the sample points to the collision boundary in the pose space. In the inference phase: given a self-penetrating pose \(\theta_0\), \(g(\theta) \ge C_l\) is embedded as a differentiable constraint into the SLSQP constrained optimizer, aiming to minimize the pose distance \(d_{\text{SMPL}}(\theta, \theta_0)\) to solve for the nearest collision-free pose \(\theta^*\). For motion sequences, the trained \(g(\theta)\) is frozen, and the input noise of the diffusion model rather than the motion itself is optimized via the DNO framework, achieving generator-agnostic temporal collision resolution.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["SMPL Pose Sampling<br/>6D Rotation Space 126D"] --> B["FCL Exact Collision Detection<br/>Binary Annotation (+1/-1)"]
    B --> C["Eikonal-Regularized Pose-Space Collision Field<br/>MLP learning g(θ): Sign Supervision + Eikonal/TD Loss"]
    C --> D["SLSQP Constrained Optimization<br/>min d_SMPL(θ,θ₀) s.t. g(θ)≥C_l"]
    D --> E["Collision-Free Pose θ*"]
    D --> F["Motion Sequence Extension<br/>DNO Framework: Optimize Input Noise x"]
    F --> G["Collision-Free Motion Sequence"]

Key Designs¶

1. Eikonal-regularized pose-space neural collision field: turning the constraint function into a pseudo-SDF

The core problem is: how to make the output of a neural network \(g(\theta)\) accurately distinguish between colliding/collision-free poses (sign supervision) while guaranteeing that the gradient norm is globally non-zero (satisfying LICQ). The authors' key insight is: for binary classification of collision/collision-free, there exists a theoretical signed distance function \(\phi(\theta)\) satisfying the Eikonal equation \(\|\nabla \phi\| = 1\) (almost everywhere), which is the unique viscosity solution with the collision boundary as its level set. Approaching this \(\phi(\theta)\) using a neural network simultaneously solves both sign accuracy and gradient regularity.

Specific implementation: Sign supervision of \(g(\theta)\) is achieved via a hinge-style loss: \(\mathcal{L}_{\text{sign}} = -\min(g(\theta) \cdot \iota(\theta), 0)\), which is zero when \(g(\theta)\) has the same sign as the true label \(\iota(\theta)\) and penalizes otherwise. Eikonal regularization \(\mathcal{L}_{\text{grad}} = |\|\nabla g(\theta)\| - 1|\) forces the gradient norm to be close to 1. The authors prove (Proposition 1) that if \(\mathcal{L}_{\text{grad}} \le \varepsilon\), the volume of the region where the Eikonal condition fails significantly (deviating from 1 by more than \(\delta\)) does not exceed \(\varepsilon/\delta\). That is, the more thorough the training, the larger the region where the LICQ holds. This proposition builds a quantitative bridge between theory and practice.

2. SLSQP constrained optimization and LICQ convergence guarantee: a theoretical closed loop from binary labels to differentiable constraints

With \(g(\theta)\) satisfying the approximate Eikonal condition, collision resolution can be formulated as a constrained optimization problem: \(\theta^* = \operatorname{argmin} d_{\text{SMPL}}(\theta, \theta_0) \text{ s.t. } g(\theta) \ge C_l\). The authors provide a complete convergence analysis based on three assumptions: (1) Smoothness: \(g\) is \(C^2\) with Lipschitz continuous gradient/Hessian (naturally satisfied by Softplus MLP); (2) Feasibility consistency: the sign of \(g\) matches the true collision label; (3) Approximate Eikonal: \(1-\delta \le \|\nabla g\| \le 1+\delta\).

Under these three assumptions, the authors prove two core theorems. Theorem 1 (Global Convergence): Since \(\|\nabla g\| \ge 1-\delta > 0\) holds globally, the LICQ is always satisfied. All limit points of a line-search SQP method (\(\ell_1\) penalty) starting from any initial point are first-order KKT points, and the iteration complexity to reach an \(\varepsilon\)-approximate KKT point is \(O(\varepsilon^{-2})\), which is of the same order as unconstrained smooth optimization—because the minimum singular value of the constraint Jacobian is bounded below by \(1-\delta\) and does not degenerate. Theorem 2 (Local Convergence): When \(\kappa = \lambda^* \|\nabla^2 g(\theta^*)\|_2 < 2\) (where \(\lambda^*\) is the KKT multiplier), the Lagrangian Hessian is positive definite, the second-order sufficient condition holds, SQP converges quadratively locally, and BFGS converges superlinearly. The condition \(\kappa < 2\) here provides a verifiable criterion for fast convergence.

The value of these theorems lies not only in proving that PoseShield is feasible, but also in revealing why Eikonal regularization is necessary: without it, vanishing gradients lead to LICQ violation and degenerate QP subproblems, causing the optimizer to stall at infeasible stationary points. This is the fundamental difference between this paper and methods like N-Penetrate that train classifiers solely using cross-entropy, which lacks gradient regularity guarantees.

3. Symmetric temporal difference training variant and active learning strategy: stable training and boundary accuracy

Directly using \(\mathcal{L}_{\text{grad}} = |\|\nabla g\| - 1|\) requires calculating second-order derivatives (the gradient of the gradient norm), which is computationally expensive and unstable in the 126-dimensional pose space. Inspired by Physics-Informed Neural Networks (PINNs) and the TD loss in offline reinforcement learning, the authors propose a symmetric temporal difference alternative: \(\mathcal{L}_{\text{TD}} = |g(\theta + v\cdot\Delta t) - g(\theta - v\cdot\Delta t) - 2\Delta t|\), where \(v = \nabla g / \|\nabla g\|\) is the normalized gradient direction, and \(\Delta t\) defaults to 0.01. Intuitively, if \(g\) strictly satisfies the Eikonal equation, the function value should change by \(\Delta t\) for each step \(\Delta t\) along the gradient direction, and the round-trip difference is exactly \(2\Delta t\). Experiments show that using \(\mathcal{L}_{\text{TD}}\) alone yields the best results (SCC 95.8%), whereas adding \(\mathcal{L}_{\text{grad}}\) degrades performance due to training instability introduced by second-order derivatives.

Furthermore, the accuracy of collision resolution heavily depends on the accuracy of \(g(\theta)\) near the collision boundary (zero level set), but random sampling rarely covers these regions. The authors adopt the active learning strategy from N-Penetrate: every 40 epochs, gradient descent is performed on a batch of samples with the objective of \(\min 0.5 |g(\theta)|^2\), and all intermediate solutions along the optimization paths are collected into the training set. These points naturally cluster near the zero level set. The entire training takes about 17 hours on a single GPU.

4. Training-free extension to motion sequences: frozen collision field + DNO framework

Collision resolution for a single pose naturally extends to temporal motion sequences. The core challenge is how to resolve collisions while maintaining motion smoothness and global structure. The authors adopt the DNO (Diffusion Noise Optimization) framework: for a colliding motion sequence \(m_s = f(x)\) generated by a diffusion model, instead of directly optimizing the motion frames \(\theta^t\), they optimize the input noise \(x\) of the diffusion model. The objective function is \(Q(m) = \sum \max(C_l - g(\theta^i), 0) + \lambda_m \cdot d_{\text{motion}}(m, m_s)\). The first term is a soft penalty version of the collision constraint (DNO does not support hard constraints), and the second term \(d_{\text{motion}}\) consists of three components: \(L_2\) distance in the pose parameter space \(\mathcal{L}_{\text{feat}}\), \(L_2\) distance of 3D joint positions \(\mathcal{L}_{\text{pos}}\), and \(L_2\) distance of joint velocities \(\mathcal{L}_{\text{vel}}\) (weights \(\lambda_{\text{joint}}=1, \lambda_{\text{vel}}=0.1\)). Together, they ensure that the corrected motion does not deviate too far from the original motion.

The elegance of this design is that \(g(\theta)\) is trained only on static poses but can be directly applied to each frame of a motion sequence without fine-tuning on motion data, achieving decoupling from the specific motion generator (such as MDM, MoMask, etc.).

Loss & Training¶

The total training objective is \(\mathcal{L}_{\text{Eikonal}} = \frac{1}{|\mathcal{D}_\theta|} \sum (\mathcal{L}_{\text{sign}}^i + \mathcal{L}_{\text{TD}}^i)\), where \(\mathcal{L}_{\text{sign}} = -\min(g(\theta)\cdot\iota, 0)\) is the sign supervision term and \(\mathcal{L}_{\text{TD}}\) is the temporal difference Eikonal loss (\(\Delta t=0.01\)). An explicit gradient norm loss \(\mathcal{L}_{\text{grad}}\) is not used, as it introduces second-order derivatives that lead to training instability. Network architecture: 12-layer MLP with a hidden dimension of 512 and Softplus activation. The optimizer used during inference is SLSQP (implemented via SciPy). The model is trained for 200 epochs, with active learning performed every 40 epochs. The constraint threshold \(C_l\) defaults to 0 and can be increased to achieve a higher collision resolution rate (at the cost of increased pose deviation MVD); the typical adjustable range is \([-0.2, 0.6]\). Single pose inference takes approximately 7.26 seconds.

Key Experimental Results¶

Main Results¶

Results on the self-built HwC dataset (931k poses, 500 subset benchmark) and the PROX dataset:

Method	HwC SCC↑	HwC PDR↑	HwC MVD↓	PROX SCC↑	PROX PDR↑	PROX MVD↓
Torch-mesh-isect	0.100	0.357	0.041	0.110	0.291	0.012
Classifier baseline	0.056	0.081	0.002	0.170	0.204	0.006
COAP	0.446	0.832	0.106	0.560	0.775	0.016
VolumetricSMPL	0.250	0.541	0.068	0.333	0.699	0.013
PoseShield (Ours)	0.958	0.982	0.059	0.800	0.893	0.021

PoseShield improves the SCC on HwC from 44.6% to 95.8% while achieving a lower MVD (0.059 vs 0.106), indicating that the corrected poses are closer to the original inputs.

Ablation Study¶

Configuration	HwC SCC↑	HwC PDR↑	HwC MVD↓	Description
Full model (L_TD + WD)	0.958	0.982	0.059	Full model
w/o Weighted Distance	0.960	0.987	0.067	MVD rises slightly; joint weights help maintain pose structure
L_grad + WD	0.862	0.917	0.062	Using only gradient norm loss; performance drops significantly
L_grad + L_TD + WD	0.870	0.922	0.062	Combining both is still worse than pure L_TD
w/o grad term	0.068	0.081	0.458	Without Eikonal regularization; almost fails completely

Motion sequence correction (100 sequences, Hymotion generator):

Method	Jitter↓	MFD↓	FSR(%)↓	RPD↓
GT (Original with Collisions)	0.598	0.000	7.47	1.721
COAP (DNO)	0.625	0.825	13.93	0.550
Direct Optimization	0.765	0.086	7.89	0.071
PoseShield (Ours)	0.514	0.401	2.42	0.017

Key Findings¶

Using \(\mathcal{L}_{\text{TD}}\) alone yields the best performance: The explicit gradient norm loss \(\mathcal{L}_{\text{grad}}\) introduces second-order derivatives leading to training instability, whereas \(\mathcal{L}_{\text{TD}}\) indirectly enforces the Eikonal property via function value differences, which is both stable and effective. Removing all gradient terms causes the model to fail completely (SCC 6.8%), proving that Eikonal regularization is the core of the method.
Weighted Joint Distance (WD) is effective: Weighting by the size of the kinematic chain subtree reduces the MVD from 0.067 to 0.059—penalizing changes in proximal joints is more reasonable than penalizing distal fingers, as torso rotation causes large displacements of the entire arm.
The value of \(g(\theta)\) strongly correlates with penetration depth: Although trained with only binary labels, the numerical value of \(g(\theta)\) (more negative indicates deeper penetration) naturally correlates positively with physical penetration depth, showing that Eikonal training implicitly encodes collision severity.
Threshold \(C_l\) provides a controllable SCC-MVD trade-off: Increasing the constraint tolerance \(C_l\) (from 0 to 0.6) improves the SCC but increases the MVD. The SCC-MVD curve of PoseShield consistently outperforms all baselines, meaning it resolves collisions more thoroughly at the same MVD.
Significant advantages on motion sequences: PoseShield's RPD (0.017) is much lower than direct optimization (0.071) and COAP (0.550), and the foot slip ratio (2.42%) is even lower than that of the original colliding sequences (7.47%), indicating that the correction process does not introduce extra foot-sliding artifacts.

Highlights & Insights¶

The theoretical bridge of Eikonal-LICQ is the most elegant design in this paper: It equates the mathematical condition for constrained optimization (non-zero gradient) with a physical geometric condition (Eikonal equation), creating a complete theoretical loop—it is not merely "Eikonal makes it work better", but rather "without Eikonal, the optimizer will mathematically inevitably fail". This paradigm of "proving a necessary condition first, then designing a loss function to satisfy it" can be transferred to any scenario requiring neural networks for differentiable constraints (e.g., collision avoidance in robot motion planning, grasp stability constraints, etc.).
Replacing gradient norm loss with TD loss: Directly regularizing the gradient norm in the high-dimensional 126D space is unstable. Implementing it indirectly via a finite-difference TD(0)-equivalent objective avoids second-order derivatives and actually yields better results. This technique, borrowed from PINNs and offline RL, has general reference value for neural field training.
A frozen static collision field directly serving motion sequences: Without needing training or fine-tuning on temporal data, it simply evaluates \(g(\theta^i)\) frame-by-frame and optimizes via DNO in the noise space of the diffusion model rather than the motion space—this preserves the generation prior's constraints on motion naturalness, representing an elegant decoupling design.
Active learning for sampling boundaries: Collecting intermediate points along the gradient descent path as samples near the boundary is a highly efficient data augmentation strategy for collision/collision-free classification tasks, where the precise shape of the collision boundary is far more important than regions far from it.

Limitations & Future Work¶

Only supports a fixed body shape: PoseShield is currently trained for a fixed \(\beta\) (body shape parameter) because the 126-dimensional pose space is already highly dimensional, and adding body shape variables would further increase the difficulty. For applications requiring generalization across body shapes (such as multi-person scenes), the network needs to be retrained or designed to condition on body shape. The authors note that in digital content creation, fixed character shapes are a common setting, making this limitation somewhat reasonable.
The distance metric is purely geometric: Both \(d_{\text{SMPL}}\) and \(d_{\text{motion}}\) are based on the \(L_2\) norm, which cannot capture semantic similarity (e.g., whether a finger indeed touches the head). In animation and interaction scenarios, semantic fidelity is often more important than geometric distance. Introducing a semantic distance metric would be a valuable extension.
Slow inference speed: Averaging 7.26 seconds per pose, it is not fast enough for real-time interactive applications (such as instant correction in VR). Although SLSQP convergence is theoretically quadratic/superlinear, it still requires multiple iterations to evaluate \(g(\theta)\) and its gradient in practice.
Soft constraint limitation of the motion sequence method: The DNO framework does not support hard constraints (e.g., \(g(\theta) \ge 0\)) and can only approximate them with soft penalties, theoretically failing to guarantee complete collision removal.
Limited size of the HwC benchmark: Testing with only 500 poses may not fully cover all collision types and severities.
No direct comparison with N-Penetrate: Because N-Penetrate is not open-sourced, the classifier baseline is an approximate reproduction rather than a direct comparison with the original.

vs COAP / VolumetricSMPL: COAP learns an occupancy field in 3D workspace and performs pose optimization by evaluating penetration loss on sampled points—essentially "checking for penetration in physical space". PoseShield directly learns a collision SDF in the pose space—essentially "checking the distance to the collision boundary in parameter space". The advantage of the latter is that the optimization variables and constraint functions reside in the same space, meaning gradients act directly on pose parameters without backpropagating through the SMPL forward kinematics chain. The performance gap (COAP's SCC 44.6% vs PoseShield's 95.8%) mainly stems from the gradient stability provided by Eikonal regularization.
vs N-Penetrate: N-Penetrate is also a pose-space classifier + optimization framework but lacks Eikonal regularization. The authors' analysis reveals why this is insufficient: without the Eikonal guarantee, the classifier's gradient can vanish at the boundary, causing the optimizer to converge to incorrect results. PoseShield's theoretical analysis elevates N-Penetrate's empirical shortcomings to a mathematical inevitability.
vs PoseNDF / NRDF: These works also learn SDF-style neural fields on the pose manifold, but their purpose is to encode the "manifold of feasible motions" (data-driven priors) for denoising and inverse kinematics. PoseShield learns specific geometric constraints (self-collision) and explicitly designs the field properties for constrained optimization rather than generation/denoising. The two are complementary—PoseShield can be embedded as a constraint term in the prior optimization of PoseNDF.
vs Classic Mesh Collision Handling (IPC, Repulsive Shells, etc.): These methods optimize in the mesh vertex space, requiring a collision-free reference configuration, and act as solvers within simulation pipelines. PoseShield operates in the pose space, is compatible with learning pipelines, requires no reference configuration, and is positioned as a post-processing module.

Rating¶

Novelty: ⭐⭐⭐⭐⭐ Formulating collision correction as a constrained optimization problem and bridging the LICQ with neural field training via Eikonal regularization is highly elegant and theoretically deep—not just a simple "adding a loss to make it better" approach.
Experimental Thoroughness: ⭐⭐⭐⭐ The main experiments cover two datasets and five baselines, with sound ablation analyses (loss terms, distance metrics, threshold trade-offs, motion extensions). However, the HwC test subset with only 500 samples is slightly small, and there is no direct comparison with the original N-Penetrate.
Writing Quality: ⭐⭐⭐⭐⭐ Clear problem definition, a complete logic chain of theory-practice-experiment, rigorous theorem statements (with full proofs in the supplementary material), and the SCC-MVD trade-off curves in the figures clearly convey the key information.
Value: ⭐⭐⭐⭐⭐ Directly practical for SMPL-related applications (pose estimation, motion generation, digital human animation). The Eikonal-LICQ design paradigm can be transferred to other scenarios where neural networks are used to replace discrete constraints.