Unified Distance Field Representations

• Unified Distance Field Representations are continuous and differentiable fields that encode spatial relationships using both analytic and learned methods.
• They unify methods like unsigned, signed, directed, and probabilistic distance fields to model shapes, collisions, and visuo-motor policies across robotics, graphics, and simulation.
• Modern techniques leverage neural networks, Gaussian processes, and advanced spatial data structures to achieve high fidelity, scalable, and task-agnostic geometric representations.

Unified distance field representations provide a mathematically principled, continuous, and often differentiable framework for encapsulating spatial relationships across computer vision, robotics, graphics, and simulation. These representations unify a multitude of geometric and physical phenomena—including shape modeling, collision avoidance, implicit surface reconstruction, and visuo-motor policy conditioning—within a shared distance field abstraction. Approaches leverage both analytic and learned models, typically grounded in Euclidean distance but increasingly generalizing to Riemannian, probabilistic, and directed distance paradigms; modern methods utilize neural networks, Gaussian process inference, and advanced spatial data structures to attain high fidelity, task-agnostic, and scalable representations.

1. Mathematical Foundations and Variants

Unified distance fields canonically encode the proximity between points and geometric loci (surface or object) via distance metrics. The fundamental structures are:

• Unsigned Distance Fields (UDFs):

$$u(\mathbf{x}) = \min_{\mathbf{y}\in S} \|\mathbf{x} - \mathbf{y}\|$$

where $S$ may be any set of surface points, enabling representation of open, closed, non-manifold, and non-orientable geometries. UDFs are agnostic to "inside" or "outside", avoiding topological constraints (Kong et al., 14 Oct 2025, Venkatesh et al., 2020).

• Signed Distance Fields (SDFs):

$$s(\mathbf{x}) = \begin{cases} +\min_{\mathbf{y}\in S} \|\mathbf{x} - \mathbf{y}\| & \mathbf{x} \text{ outside } S\ -\min_{\mathbf{y}\in S} \|\mathbf{x} - \mathbf{y}\| & \mathbf{x} \text{ inside } S \end{cases}$$

SDFs encode both magnitude and occupancy, but require consistent orientation and watertightness.

• Directed Distance Fields (DDFs) and Probabilistic DDFs (PDDFs):

$d(p,v) = \min\{ t \ge 0 : p + t v \in S \}$

for a point $p$ and direction $v$, optionally augmented with a visibility indicator $\xi(p, v)$, and probabilistic mode mixtures to capture discontinuities (Aumentado-Armstrong et al., 2024, Aumentado-Armstrong et al., 2021).

• Configuration Space Distance Fields (CSSDFs):

Joint-space extensions modeling collision safety in configuration space, fusing environment and self-collision in a single continuous field (Chen et al., 19 Mar 2026).

• Bernstein Polynomial Robot Distance Fields (RDF):

Flexible, smooth SDF representations for articulated bodies via tensor-product Bernstein polynomials and kinematic composition (Li et al., 2023).

• Probabilistic Distance Fields (GPDF):

The Euclidean distance (or log-distance) field modeled as the mean of a Gaussian Process, enabling uncertainty-aware, continuous inference and gradients (Wu, 2024, Wu et al., 2024).

• Factor Field Decompositions:

Unified frameworks (e.g., Dictionary Fields/DiF) decomposing signals into products of neural and classical fields, encompassing SDFs, occupancy nets, radiance fields, and hash-grid approaches (Chen et al., 2023).

2. Construction Methodologies and Algorithms

Construction approaches vary by scale, fidelity, and type of domain (scene, articulated object, configuration space):

• Analytic and Polynomial Encodings: Bernstein polynomials for fast, differentiable per-link SDFs, composed via kinematics to represent full-body robot geometry, enabling analytic derivatives and rapid gradient queries (Li et al., 2023).
• Neural Implicit Functions: Multi-layer perceptrons (MLPs) fit to sparse or dense surface samples, either as SDFs (DeepSDF), UDFs/normals (DUDE), or hybrid neural fields (Factor Fields, DiF). Positional encoding, residual connections, and architectural regularization are used for detail preservation and expressiveness (Venkatesh et al., 2020, Chen et al., 2023).
• Gaussian Process Inference: Local or global GP regression over spatial samples (surface, occupancy, or zero-crossing) delivers continuous probabilistic fields; variants include log-GPIS and reverting transformations for exact Euclidean correspondence. Efficient scaling leverages spatial partitioning and submapping (Wu, 2024, Wu et al., 2024).
• Hybrid Data Structures: Integration of GPs with sparse volumetric structures (OpenVDB) supports scalable, real-time mapping and meshing (VDB-GPDF), fusing predictions via weighted uncertainty-aware updates (Wu et al., 2024).
• Voronoi-Assisted Diffusion: Network-free methods infer UDFs from unoriented point clouds via Voronoi-based bi-directional normal alignment and tensor diffusion, followed by Poisson integration for the scalar field. These preserve UDF properties even for challenging topologies (Kong et al., 14 Oct 2025).

3. Unification Across Application Domains

Unified distance field representations underpin geometric and physical computation for a spectrum of domains:

• Robotics: Unified CSSDFs enable differentiable joint-space safety margins for manipulators, supporting both motion planning (offline spline optimization with analytic gradients) and online receding-horizon MPC with linearized safety constraints, validated on high-DoF systems and dynamic obstacles (Chen et al., 19 Mar 2026). Robot Distance Fields allow efficient, smooth collision and contact queries in whole-body planning and manipulation (Li et al., 2023).
• 3D Vision and Graphics: Factor Fields (including DiF-Grid) subsume a range of geometric neural representations, delivering high geometric IoU and training efficiency, with direct application to image regression, few-shot scene modeling, and radiance fields (Chen et al., 2023). DUDE unifies surface geometry and normals for open/closed shape modeling; VAD UDFs extend unified implicit modeling to arbitrary topology, outperforming prior methods in error and robustness (Venkatesh et al., 2020, Kong et al., 14 Oct 2025).
• Perception-to-Control Pipelines: Distance field cues (distance, gradient, surface-relative velocities) drive learned interaction policies in humanoid robotics, supporting scale and geometry generalization, long-horizon task composition, and vision-only transfer (LessMimic) (Lin et al., 25 Feb 2026).
• Dense Mapping and Reconstruction: Probabilistic and VDB-GPDF pipelines perform efficient, uncertainty-aware, dense surface and distance field mapping, supporting direct downstream queries for planning and visualization with frame-level latencies (Wu et al., 2024, Wu, 2024).
• Differentiable Rendering and Inverse Graphics: DDFs and PDDFs permit single-pass rendering of depth and geometry with full backpropagation support, view consistency guarantees, fast compositionality, and native support for handling discontinuities and occlusion events (Aumentado-Armstrong et al., 2024, Aumentado-Armstrong et al., 2021).

4. Advanced Properties and Unified Field Benefits

Unified distance field frameworks deliver several pivotal mathematical and practical properties:

• Continuity and Differentiability: Analytic constructions (e.g., Bernstein polynomials), neural MLPs, and GP-based methods yield fields that are at least piecewise smooth, often $C^\infty$, enabling reliable computation of gradients and higher derivatives. This is critical for motion planning, surface normal extraction, and optimization-based tasks (Li et al., 2023, Wu, 2024).
• Unified Treatment of Geometry and Safety: CSSDF-Net and RDF unify self- and environment collision as a single field, providing a coherent safety margin and robust, stable gradients directly usable in constrained optimization (Chen et al., 19 Mar 2026, Li et al., 2023).
• Probabilistic Uncertainty Quantification: GP-based fields (GPDF, VDB-GPDF) furnish predictive variance, informing the reliability of queries and facilitating uncertainty-aware control, exploration, and fusion of heterogeneous sensor data (Wu, 2024, Wu et al., 2024).
• Topology and Robustness Agnosticism: UDF and DDF-based approaches, notably DUDE and VAD, eliminate reliance on watertight input and are robust to open, non-manifold, self-intersecting, or non-orientable domains, which underpins their flexibility (Venkatesh et al., 2020, Kong et al., 14 Oct 2025).
• Multi-task and Multi-scene Generalizability: Unified field architectures (Factor Fields, LessMimic) allow for training and inference across mixed geometric classes, scenes, and tasks without retraining or pose-specific engineering, supporting transfer and composition in both vision and control (Chen et al., 2023, Lin et al., 25 Feb 2026).

5. Comparative Performance, Scalability, and Limitations

Quantitative and empirical evaluations consistently demonstrate the practical viability and performance edge of unified distance field approaches:

• High-fidelity geometric modeling (gIoU $>$0.9 for DiF-Grid, superior Chamfer/Hausdorff Distance for VAD and DUDE).
• Low-latency, real-time inference: CSSDF-Net achieves sub-5 ms for $10^4$ queries on GPU, enabling online MPC at $\sim$280 Hz (Chen et al., 19 Mar 2026).
• Scalability: VDB-GPDF matches or exceeds prior volumetric baselines in accuracy and is competitive in memory and runtime due to hierarchical data management and local-GP fusion (Wu et al., 2024).
• Versatility: DDF/PDDF frameworks demonstrate rapid single-image 3D reconstruction (sub-10 ms per render), direct extraction of normals/curvatures, and efficient backpropagation for learning tasks (Aumentado-Armstrong et al., 2024).

Limitations remain in areas including:

• Scalability of GP-based approaches for large $N$ without submapping or inducing-point approximations (Wu, 2024).
• Inference overhead or memory consumption for high-resolution, full-scene neural fields, especially if global rather than factorized (Chen et al., 2023).
• Sensitivity to topological complexity in SDF-based pipelines, if not explicitly unified as in UDF, DDF, or hybrid approaches (Venkatesh et al., 2020, Kong et al., 14 Oct 2025).

6. Theoretical Guarantees and Field Consistency

Unified distance field representations have been rigorously analyzed for shape induction, consistency, and field-theoretic properties:

• View Consistency in Directed Fields: DDFs/PDDFs admit a complete local field characterization guaranteeing the existence of a unique underlying shape, provided certain boundary, monotonicity, and isotropy constraints are met (see Theorems: Simple DDF⇔Shape, Visibility field⇔Shape Indicator, Full DDF⇔Shape Representation) (Aumentado-Armstrong et al., 2024).
• Eikonal Regularity: Many field constructions (CSSDFs, DUDE, VAD) explicitly enforce Eikonal constraints ($\|\nabla u\| = 1$ almost everywhere), ensuring metric validity for gradient-based planning and optimization (Chen et al., 19 Mar 2026, Kong et al., 14 Oct 2025).
• Loss Decomposition and Training Objectives: Modern unified representations employ composite losses (distance, Eikonal, direction/cosine) to stabilize and specialize the learned field; mixture weight variance and transition losses regulate abrupt transitions critical for capturing geometric discontinuities (Chen et al., 19 Mar 2026, Aumentado-Armstrong et al., 2024, Venkatesh et al., 2020).

These advances underpin the reliability of unified distance fields as a foundation for a broad array of geometric reasoning and embodied intelligence algorithms.