Euclidean Signed Distance Field (ESDF)

• ESDF is a scalar field that assigns signed Euclidean distances to the nearest surface, using positive values for free space and negative values inside obstacles.
• It employs methods ranging from explicit grid-based algorithms and GPU acceleration to neural representations, ensuring continuous and differentiable distance queries.
• ESDF enables efficient collision checking, trajectory optimization, and real-time scene reconstruction, which are critical for autonomous systems and robotics applications.

Euclidean Signed Distance Field (ESDF) is a scalar field defined over a spatial domain, commonly ℝ³, that encodes at each point the signed Euclidean distance to the nearest surface within a set of obstacles or observed surfaces. By convention, positive values represent free space, negative values indicate interior to obstacles, and the zero-level set corresponds to the surface boundary. ESDFs are foundational in motion planning, mapping, surface reconstruction, and scene understanding due to their continuous, differentiable, and physically meaningful geometric properties. They support precise collision checking, trajectory optimization, and enable downstream reasoning tasks that demand efficient spatial queries and distance gradients.

1. Mathematical Formalism and Core Properties

The ESDF is formally defined as follows. Let $\Omega_{\rm obs} \subset \mathbb{R}^3$ denote the set of observed surface points. For any $x \in \mathbb{R}^3$: $S(x) = \pm \min_{y \in \Omega_{\rm obs}} \|x - y\|$ where the sign convention is $+1$ for $x$ in free space, $-1$ for $x$ inside an obstacle. The field $S: \mathbb{R}^3 \rightarrow \mathbb{R}$ thus satisfies:

• Continuity everywhere
• Differentiability almost everywhere, with $|\nabla S(x)| = 1$ except at medial axes
• The gradient $\nabla S(x)$ points towards the nearest surface point
• The zero-level set $\{x : S(x) = 0\}$ is the reconstructed surface

This formulation underpins nearly all contemporary mapping and optimization applications involving signed distance computations (Yue et al., 2024).

2. Classical and Efficient ESDF Construction Algorithms

Efficient computation of the ESDF is central to enabling real-time operation in autonomous systems. Traditional methods include:

Explicit Grid-Based Incrementation

• BFS-based updates: Methods such as FIESTA maintain an occupancy grid and propagate true Euclidean distances using BFS when occupancy flips (insertions/deletions). Data structures such as doubly linked lists and hash tables for index mapping allow O(1) lookup and efficient local updates. Accuracy depends on connectivity; 24-connectivity achieves RMS error ≈0.2 voxels in <20 ms at 0.05 m resolution (Han et al., 2019).
• TSDF-to-ESDF Lifting: Voxblox generates a TSDF layer by fusing sensor data and then incrementally builds an ESDF by propagating values from a “fixed band” of anchored voxels using dual (raise/lower) wavefront queues. This architecture runs at 4–10 Hz for large scenes on a CPU, with ESDF computation times <10 ms per frame (Oleynikova et al., 2016).

GPU Acceleration

• nvblox delivers a fully parallel, block-sweeping ESDF update on the GPU. It leverages blockwise hash grids and propagates true Euclidean distances from TSDF zero-crossings. Achieves up to 31× speed-up over CPU methods, with per-frame ESDF fusion times of 1.9 ms at 0.05 m voxel resolution (Millane et al., 2023).
• OpenVDB and Octree Approaches: VDB-based systems support fast access and hierarchical memory savings (30–85% over octrees in sparse scenes), but complex queries may incur higher costs in large scenarios (Wu et al., 2024).

3. Implicit, Neural, and Probabilistic Representations

Implicit and neural architectures bypass fixed grids, offering fully continuous ESDFs:

• MLP-Based Networks: LGSDF trains a 6-layer MLP to regress $f_\theta: \mathbb{R}^3 \rightarrow \mathbb{R}$ using local, fused grid observations for supervision, with additional gradient and Eikonal penalties. Continual learning is achieved by mixing “current” and “historical” grid cell samples per batch, preserving both recent plasticity and global stability. In ReplicaCAD and ScanNet, LGSDF reduced mean absolute SDF error to 2.95 cm and mesh completion error to 9.82 cm, outperforming prior explicit and implicit methods (Yue et al., 2024).
• Hybrid Explicit–Implicit Hierarchies: $\nabla$-SDF combines an explicit, gradient-augmented octree prior with an implicit neural residual component, supporting non-truncated, globally differentiable ESDFs with a memory footprint comparable to sparse TSDF grids. Online updates and stochastic training integrate new measurements incrementally, achieving valid ESDF coverage over 100% of volume and gradient MAE ≈0.33 rad (Dai et al., 21 Oct 2025).
• Gaussian Process Distance Fields: VDB-GPDF leverages GP-based local clusters fused hierarchically in OpenVDB. At each voxel, the distance and its uncertainty are inferred from local GP posterior means; global fusion uses information-weighted updates. This method achieves ESDF RMSE ≈0.02 m (vs. 0.06 m for Voxblox) at 0.05 m voxels (Wu et al., 2024).

4. Variational and Elliptic PDE Approaches

Spectral and PDE-based methods present alternative strategies:

• Elliptic Equation Formulation: The elliptic-PDE approach solves $-\alpha \Delta u + u = f$ in a domain $\Omega$, encoding the set $A$ of interest by $f$ and extracting the ESDF as $U_\alpha(x) = \pm \sqrt{\alpha} \log u_\alpha(x)$. Uniform convergence to the true distance occurs as $\alpha \to 0^+$, with proven $O(\sqrt{\alpha}\log(1/\alpha))$ error in 1D (Hasebe et al., 2024).
• Convolutional FFT-based Algorithms: Gurumoorthy & Rangarajan formulate a variational principle yielding $S = -\tau \log \phi$ where $\phi$ satisfies a linear PDE; closed-form solutions are computable with FFT at $O(N\log N)$, with $O(\tau\log K)$ error. Sign determination is performed via winding number in 2D or topological degree in 3D (Gurumoorthy et al., 2011).

5. Learning Objectives and Regularization

Modern ESDF learning frameworks employ composite loss functions:

• Supervised and Self-Supervised Terms: SDF fitting, gradient matching, and zero-level boundary constraints form the core, as in LGSDF: $$\mathcal{L} = \sum_v L_{\rm sdf}(f_\theta(v), D(v)) + \lambda_{\rm grad}\sum_v L_{\rm grad}(\nabla f_\theta(v), G(v)) + \lambda_{\rm eik}\sum_v L_{\rm eik}(\nabla f_\theta(v))$$.
• Eikonal Constraints: Enforced via $|\nabla \phi(x)| - 1$ penalties, ensuring the field approximates a true Euclidean metric, as in SurroundSDF’s sandwich formulation and implicit neural SDF techniques (Liu et al., 2024, Fayolle, 2021).
• Gradient and Hermite Interpolation: Hybrid methods (e.g. $\nabla$-SDF) explicitly store gradients at octree nodes, allowing Hermite interpolation and low reconstruction error.

6. Applications, Performance, and Limitations

ESDFs are the backbone of:

• Motion Planning and Trajectory Optimization: ESDFs enable differentiable collision penalties and support analytic path gradient computation, essential for real-time MPC and trajectory planning (e.g., CHOMP, C-3TO) (Gil et al., 24 Sep 2025, Oleynikova et al., 2016).
• Online Mapping and Surface Reconstruction: In robotics, systems such as LGSDF and FIESTA maintain and update the ESDF in parallel with state estimation.
• Semantics and 3D Scene Understanding: SurroundSDF combines per-point ESDF prediction with semantic segmentation, supervised by LiDAR and weak geometric constraints for robust scene reconstruction (Liu et al., 2024).
• Visibility and Occlusion Reasoning: Specialized, precomputed structures such as FoV-ESDF in Eva-Tracker enable efficient, update-free visibility-aware trajectory planning with low computational overhead (Lin et al., 13 Feb 2026).

Practically, ESDF representation choice involves trade-offs among memory, computational load, update latency, spatial resolution, and query/readout speed. While neural and hybrid methods now approach or exceed explicit approaches in accuracy and gradient quality, limitations persist for large-scale, dynamic environments (MLP scalability, latency), as well as for very high update rates.

7. Future Directions and Open Challenges

Scalability to city-scale or rapidly dynamic environments is an open challenge, with promising directions including:

• Hierarchical and Multi-Resolution Structures: Integrating hierarchical grids, multi-resolution hashmaps, or dynamic adaptation for efficient memory scaling (Yue et al., 2024).
• Probabilistic and Uncertainty-Aware ESDFs: Incorporation of uncertainty at each node (e.g., VDB-GPDF) and probabilistic fusion for robust mapping under noisy observations (Wu et al., 2024).
• Semantic and Multi-Modal Fusion: Joint modeling of distance and semantics, as in SurroundSDF, supports more robust and informative scene understanding.
• Direct Integration in Trajectory Optimization: Online, differentiable ESDFs can enable real-time, continuous replanning for aerial robotics and autonomous systems (Gil et al., 24 Sep 2025).
• Hardware Acceleration: Exploiting GPU and FPGA architectures for further gains in update and query throughput (Millane et al., 2023).

Future work will likely address improved handling of non-stationary scenes, fusing multi-sensor data, and more sophisticated learning objectives to further close the gap between explicit geometric accuracy and the expressive power of neural and probabilistic models.