Gaussian Process Distance Field (GPDF)

• GPDF is a probabilistic representation that uses Gaussian Process regression to generate continuous distance fields with analytically accessible derivatives.
• It unifies multiple robotics tasks by providing a single, queryable field for localisation, mapping, planning, obstacle avoidance, and grasping.
• The approach employs kernel inversion, logarithmic transformations, and uncertainty proxies to overcome traditional occupancy and surface reconstruction limitations.

Searching arXiv for recent and foundational papers on Gaussian Process Distance Field. Gaussian Process Distance Field (GPDF) denotes a family of Gaussian-process-based probabilistic distance-field representations in which a latent GP over space is queried continuously to recover distance to surfaces together with derivatives and uncertainty. In robotics, GPDF is presented as a unified representation that mathematically models the fundamental property of the Euclidean distance field (EDF) along with gradients, surface normals, and dense reconstruction, so that localisation, mapping, motion planning, obstacle avoidance, grasping, human-robot collaboration, and dense visualization can query the same field rather than rely on separate task-specific maps (Wu, 2024). The term covers closely related formulations, including log-transformed GP implicit surfaces for faithful EDF recovery (Wu et al., 2020), kernel-inverse or reverting-function methods for more accurate distance estimation (Gentil et al., 2023), online volumetric variants coupled to OpenVDB (Wu et al., 2024), and application-specific extensions for dynamic grasping (Choi et al., 2023), active visual-tactile exploration (Choi et al., 7 Jul 2025), and safe navigation with obstacle and ground constraints (Uttsha et al., 2024).

1. Conceptual scope and representational role

GPDF emerged from a robotics representational problem: different subsystems tend to use different maps, even though those maps are all partial views of the same geometry. In that framing, feature-based maps support SLAM localisation and sparse mapping but are poor for dense reconstruction, planning, and interaction; raw point clouds carry geometry but are memory-heavy and low-level; volumetric fields such as TSDF and ESDF support planning and manipulation but can be expensive to compute globally and are not as naturally suited to localisation from sparse correspondences; and neural implicit fields are described as “black-box” representations lacking theoretical transparency. GPDF is positioned as the alternative in which continuity, probabilistic uncertainty, and theoretical grounding are all explicit properties of the map itself (Wu, 2024).

The representational target is not merely an implicit surface. In the GPDF literature, the central object is a distance field that can be queried at arbitrary spatial locations and differentiated analytically. That distinction is important because many downstream robotic computations are inherently differential: optimisation-based planning requires collision costs and their gradients, grasp synthesis requires surface normals, dense visualisation requires a reconstructable zero level set, and probabilistic fusion benefits from uncertainty-aware estimates. The representation is therefore used less as a static reconstruction artifact than as a queryable geometric substrate.

The term is also used with some variation across papers. Some formulations aim at an unsigned Euclidean distance field, especially when the derivation is tied to the Eikonal equation or to global distance estimation. Other formulations emphasise approximate signed distance fields, especially when grasping, rendering, or contact reasoning require orientation relative to the surface. This variation reflects differences in kernel choice, transformation, and sensing setup rather than a single fixed canonical model.

2. Mathematical formulation

At the foundation of GPDF is standard Gaussian Process regression over a latent spatial function,

$$f(\mathbf{x}) \sim \mathcal{GP}(m(\mathbf{x}), k(\mathbf{x},\mathbf{x}')),$$

with predictive distribution at a query point $\mathbf{x}_*$,

$$p(f_* \mid \mathbf{x}_*, \mathbf{X}, \mathbf{y}) = \mathcal{N}(\mu_*, \sigma_*^2).$$

In many GPDF constructions, the latent field is occupancy-like rather than distance-like. A common formulation writes

$$o(\mathbf{x}) \sim \mathcal{GP}(0,k_o(\mathbf{x},\mathbf{x}')),$$

and infers the posterior mean from surface observations $\mathbf{X}$ with labels $\mathbf{1}$,

$$\hat{o}(\mathbf{x}) = k_o(\mathbf{x},\mathbf{X})\big(K_o(\mathbf{X},\mathbf{X})+\sigma_o^2 I\big)^{-1}\mathbf{1}.$$

Distance is then recovered by a kernel-specific inverse or reverting function $r$, so that

$\hat{d}(\mathbf{x}) = r(\hat{o}(\mathbf{x})),$

with the defining property

$$r\big(k_o(\mathbf{x},\mathbf{x}')\big)=\|\mathbf{x}-\mathbf{x}'\|=d.$$

This is the central construction in the accurate GP-based distance-field formulation, where the latent GP behaves as a smooth occupancy map and the reverting step “undoes” the kernel transform to produce a more faithful Euclidean distance estimate (Gentil et al., 2023).

A complementary theoretical route derives the field from the regularised Eikonal equation. The Euclidean distance field satisfies

$\mathbf{x}_*$0

but that PDE is nonlinear. Log-GPIS addresses this by introducing a latent heat-like field $\mathbf{x}_*$1 and the logarithmic transform

$\mathbf{x}_*$2

using Varadhan’s formula to approximate the EDF through the logarithm of a linear PDE solution. In that formulation, members of the Matérn covariance family directly satisfy the linear PDE, which yields a GP prior aligned with distance-field geometry and removes the need for post-processing to recover the EDF (Wu et al., 2020).

Derivatives are an essential part of the representation. Because derivatives of a GP are again GPs, the gradient is available by differentiating the kernel. In the reverting-function formulation, the distance gradient is written as

$\mathbf{x}_*$3

Surface normals follow conceptually from the distance gradient,

$\mathbf{x}_*$4

with normalisation giving the unit normal. Recent active-perception work extends this differential structure further by deriving closed-form Hessians and using them for mean and Gaussian curvature estimates, making GPDF useful not only for normals but also for local differential geometry and curvature-aware exploration (Choi et al., 7 Jul 2025).

Uncertainty is likewise intrinsic to the GP model, but its interpretation depends on the transformation. The latent occupancy variance is standard GP posterior variance. After nonlinear transformations such as logarithmic recovery or kernel inversion, distance variance becomes harder to model reliably. This is why later work introduced a proxy uncertainty for distance estimates rather than treating transformed variance as exact (Gentil et al., 2023). The difficulty of variance prediction under nonlinear distance recovery is a recurring limitation in the literature (Wu, 2024).

3. Historical development and relation to adjacent methods

The immediate lineage of GPDF runs through Gaussian Process Implicit Surfaces (GPIS). Standard GPIS reconstructs an implicit surface near observations, but its scalar values degrade as distance estimates away from the surface because the GP mean regresses toward its prior. Log-GPIS was introduced precisely to correct that failure mode: by applying a logarithmic transformation to a GPIS formulation and exploiting Varadhan’s formula, it recovers a field that is much closer to a faithful Euclidean distance field everywhere in space (Wu et al., 2020).

The next major step was the reverting-function formulation, which made the latent field explicitly occupancy-like and distance recovery explicitly kernel-specific. In that view, the kernel itself induces a monotone map from Euclidean distance to latent field value, so distance can be recovered by inversion. The 2023 formulation presented this as a way to overcome both GPIS and LogGPIS limitations, including the decay-to-zero behaviour of GPIS and the accuracy–interpolation trade-off of LogGPIS, and also introduced a proxy for uncertainty suited to the transformed field (Gentil et al., 2023).

Online scalability then became the central problem. VDB-GPDF addresses this by combining local GP distance fields and a global VDB hierarchy. A temporary local model, the Local Gaussian Process Signed Distance Field (L-GPDF), performs frame-wise inference in local VDB leaf nodes; uncertainty-weighted fusion integrates those predictions into a global VDB; and a Global Gaussian Process Signed Distance Field (G-GPDF) is then trained on zero-crossing points to provide accurate global distance and gradient queries (Wu et al., 2024). This preserves the continuous and probabilistic properties of GP distance fields while addressing the cost of global GP inference.

The boundary of the concept is also significant. Not every Gaussian-based distance representation is a GPDF in the strict sense. SplatlessDF, for example, uses anisotropic Gaussian primitives whose weighted sum directly parameterises a differentiable distance field; it is explicitly contrasted with GPDF because it does not use GP inference, kernel-matrix operations, or pseudo-input optimisation (Uttsha et al., 12 Jun 2026). Likewise, Gaussian surface models and Gaussian mixture models can support continuous distance, gradient, and collision-probability queries, but those methods rely on ellipsoid geometry rather than GP regression (Goel et al., 2024). By contrast, a 2026 mixture-of-GPs shape representation remains closer to the GPDF family, but it is directional and patch-wise rather than a single global Euclidean distance field: local GP priors are anchored at interior reference points and combined as a mixture to model complex topology (Sapoutzoglou et al., 1 Apr 2026).

A common misconception is therefore that “Gaussian” in the representation name refers generically to any Gaussian primitive or Gaussian surface model. In the GPDF literature proper, the defining mechanism is GP regression over a latent field, together with a mathematically specified transformation or inverse map that recovers distance and its derivatives.

4. Robotics functions and task integrations

The unifying claim for GPDF is that one field can support a broad robotics stack. The survey formulation explicitly lists localisation, mapping, motion planning, obstacle avoidance, grasping, human-robot collaboration, and dense visualization. For localisation, accurately inferred distance fields allow pose estimation via direct optimisation of alignment between the global field and new measurements. For mapping, GPDF produces a continuous and probabilistic dense surface. For planning and obstacle avoidance, the field supplies both distance and gradient, which makes collision checking and optimisation-based trajectory improvement directly queryable. For grasping, surface normals derived from the field support contact reasoning. For visualisation, the field can be converted into a densely reconstructed mesh (Wu, 2024).

Dynamic grasping is one of the clearest demonstrations of this multi-role use. In that setting, GPDF reconstructs an object’s approximate shape from a partial point cloud, provides an approximate signed distance field with gradients and uncertainty, and supports iterative refinement inspired by sphere marching or ray marching. The reconstructed surface points and GPDF-derived normals are fed into an SE(3)-equivariant grasp sampler; candidate grasps are checked against the field to ensure gripper points lie outside the object surface; and contact-pair normals are used for antipodal validation. The same paper further repurposes GPDF as a potential field in task space or joint space for reactive control. In static-object simulation on 50 YCB objects, the full pipeline with grasp sampler, post-filtering, and reachability analysis reaches a 90.00% success rate over 500 trials, and the reported ablation shows that GPDF-based post-filtering is necessary for robust grasp selection (Choi et al., 2023).

Active perception extends the representation from geometry estimation to closed-loop sensing. A recent visual-tactile framework uses GPDF to fuse RGB-D point clouds and tactile observations into a single signed-distance representation that exposes analytic gradient, Hessian, surface uncertainty, and potentially non-geometric attributes. Vision updates are incorporated through differentiable rendering, while tactile contacts are directed toward uncertain surface regions by maximising surface uncertainty under a contact constraint. The same work argues that the GP structure can, in principle, store RGB colour, CLIP embeddings, force fields, and potentially material properties by changing the observation values while keeping the spatial GP machinery fixed (Choi et al., 7 Jul 2025).

Navigation uses a different decomposition. A dual-field method formulates a Ground GPDF from segmented ground points and an Obstacle GPDF from height-relevant non-ground points projected into the $\mathbf{x}_*$5-$\mathbf{x}_*$6 plane, then incorporates both fields into a CHOMP-style optimiser. The obstacle field enforces collision clearance, while the ground field keeps the trajectory at the desired height above uneven terrain. This separation is designed for ground-based mobile systems, including wheeled robots, legged robots, and human assistance, where overhanging obstacles and steps make the problem intrinsically three-dimensional (Uttsha et al., 2024).

5. Computational organisation and online implementations

The principal computational obstacle for GPDF is standard GP scaling. Exact GP inference is cubic in the number of samples, which becomes prohibitive for online mapping, long-term operation, or dense multi-sensor fusion. This issue is explicitly identified as a major challenge for online use and one of the main reasons GPDF required architectural refinements beyond the original formulations (Wu, 2024).

VDB-GPDF is the best-defined online architecture in the current literature. A depth sensor produces a point cloud that is transformed into the world frame and voxelised into a local hierarchical VDB. Each VDB leaf node acts as a GP cluster, and together these local GPs form a temporary Local Gaussian Process Signed Distance Field. The system then generates a reduced set of testing points by ray-casting through the local VDB and augmenting near-surface coverage along estimated normal directions. For each testing point, the closest local GP node returns distance, surface property, and variances. These inferred quantities are fused into a global VDB by uncertainty-weighted averaging, after which marching cubes recovers a mesh and zero-crossing points are used to train a global GP model. To reduce discontinuities across node boundaries, the global query stage uses a smooth minimum across nearby GP nodes (Wu et al., 2024).

Other scalability strategies are more approximate. The active exploration framework discusses inducing point methods, Nyström approximation, SKI / KISS-GP, Hilbert-space approximation, voxel downsampling with weighted covariance statistics, and incremental block-matrix inverse updates for online insertion and deletion. In that account, Nyström is preferred when smoothness and out-of-bounds generalisation matter, while SKI is the fastest but may degrade gradient continuity. The same work retains the original GP for smaller problems and uses Nyström for larger ones, which reflects a broader pattern in GPDF research: exact inference remains the geometric reference point, but practical systems increasingly rely on structured approximation (Choi et al., 7 Jul 2025).

The online mapping literature also preserves the task-driven query philosophy of GPDF. Unlike global TSDF or ESDF propagation, which compute distance fields over an entire discretised volume, GP-based methods can compute distances and derivatives on demand, only where a downstream module queries the field. This makes the representation particularly attractive when the set of task-relevant queries is sparse relative to the size of the environment (Wu, 2024).

6. Limitations, interpretive issues, and future directions

The literature is explicit that GPDF is not a universally solved representation problem. One reported limitation is a trade-off between accuracy and interpolation ability in the original formulations. Another is that highly accurate Euclidean distances remain difficult in applications such as echolocation for robotic inspection, which partly motivated the reverting-function formulation. Nonlinear transformations complicate variance modelling, so uncertainty is often best interpreted near the surface rather than far from observations. More generally, the GP core still carries cubic computational complexity, current systems are often spatial-only rather than temporal or semantic, and several implementations assume depth-sensor-based input streams (Wu, 2024).

A second interpretive issue is sign. Some formulations are fundamentally unsigned, especially those derived by stitching inside and outside distance problems into a globally valid Euclidean field or those intended primarily for navigation and reconstruction. Other formulations emphasise signed distance, particularly when grasp validation, rendering, or contact reasoning requires determining whether query points lie inside or outside the surface. In some unsigned pipelines, reconstruction requires a separate sign-recovery step before marching cubes; in others, sign is recovered from sensor geometry or gradient direction. This does not amount to a contradiction, but it does mean that “GPDF” names a methodological family rather than one invariant mathematical object (Wu et al., 2020).

A third issue concerns what should count as a unified representation. The 2024 overview presents GPDF as a “one for all” representation for multiple robotics tasks (Wu, 2024). At the same time, application papers often augment it with additional structures: SE(3)-equivariant networks and Riemannian Mixture Models in dynamic grasping (Choi et al., 2023), differentiable rendering and tactile planning machinery in active exploration (Choi et al., 7 Jul 2025), or quadtrees and connectivity graphs in navigation (Uttsha et al., 2024). This suggests that GPDF unifies core geometric querying rather than eliminating the need for task-specific optimisation layers.

The stated future directions are correspondingly pragmatic. They include online GPDF with real-time outputs of distance, gradients, repulsive information, reconstruction, and colour; improved efficiency for large-scale and long-term mapping; tighter integration with a new motion planner using all GPDF outputs; extension to human-robot collaboration and manipulation; better support for dynamic environments; combination of feature-based odometry with GPDF-based odometry; expansion from purely spatial modelling to temporal information and scene understanding; potential integration with Gaussian Splatting for high-quality reconstruction; and future multi-sensor probabilistic data fusion (Wu, 2024). Taken together, these directions indicate that the central research question has shifted from whether GP distance fields can represent geometry to how far one probabilistic distance-field substrate can be extended across perception, planning, interaction, and multimodal scene understanding.