Evoxels: Enriched Voxel Representations

• Evoxels are attribute-bearing voxel representations that integrate geometric, semantic, and physical data to enhance computational applications.
• They aggregate sensor inputs and simulation parameters to support robust event recognition, robotics mapping, and materials modeling.
• Evoxel methods balance dense spatial structures with task-aware augmentations, yielding efficiency gains and improved accuracy across disciplines.

In current research usage, evoxels does not denote a single canonical data structure. The term is used for several voxel-centered representations in which each voxel carries information beyond binary occupancy: motion-sensitive aggregates of event-camera measurements, traversability statistics for robotics, semantic feature histories for CAD/CAE, graph-theoretic descriptors of free volume in atomistic systems, differentiable field variables for microstructure simulation, and per-voxel mechanical property fields for simulation-ready 3D assets (Xie et al., 2023, Overbye et al., 2021, Li et al., 2023, Chapman et al., 2021, Daubner et al., 29 Jul 2025, Dagli et al., 27 Oct 2025). A unifying view is that an evoxel is a voxel enriched with task-specific state, together with operators that make that state computationally useful.

1. Terminological scope and shared abstraction

Across the literature, the common abstraction is a regular or sparsified volumetric discretization whose cells are not merely marked “occupied” or “empty,” but are assigned geometric, semantic, physical, or topological attributes. The exact content of an evoxel depends on the application domain, and the same word therefore names distinct but structurally related ideas.

Domain	Voxel enrichment	Representative source
Event-camera recognition	Feature vector $\mathbf{f}_i$ plus spatiotemporal coordinate $\mathbf{c}_i$ in a sparse voxel set $\mathcal{V}=\{(\mathbf{f}_i,\mathbf{c}_i)\}$	(Xie et al., 2023)
Event-camera graph learning	Voxel-wise vertices with 2D patch features and voxel coordinates $\mathcal{U}_i$	(Deng et al., 2021)
Off-road mapping	Hit count $N_h$, miss count $N_m$, and lowest return height $z_{\min}$, with derived obstacle, slope, and roughness layers	(Overbye et al., 2021)
Atomistic free-volume analysis	Voxel center, maximal free-space sphere radius $r_k$, local volume, and graph degree	(Chapman et al., 2021)
CAD/CAE integration	Lists of $$(\text{feature index},\text{feature nature},\text{occupancy completeness})$$ per voxel	(Li et al., 2023)
Differentiable microstructure simulation	Voxelized scalar/vector fields managed by VoxelFields and VoxelGrid	(Daubner et al., 29 Jul 2025)
Volumetric mechanics prediction	Per-voxel Young’s modulus $E$, Poisson’s ratio $\mathbf{c}_i$0, and density $\mathbf{c}_i$1	(Dagli et al., 27 Oct 2025)

This diversity suggests that “evoxels” is best understood as a family of attribute-bearing voxel representations. A common misconception is to equate evoxels with plain occupancy grids. The surveyed systems instead use voxels as carriers of local descriptors, physical parameters, feature semantics, or graph structure, and the downstream algorithms are designed around those enrichments rather than around occupancy alone.

2. Event-camera evoxels

In neuromorphic vision, evoxels arise from the discretization of an asynchronous event stream

$\mathbf{c}_i$2

where $\mathbf{c}_i$3 is the pixel location, $\mathbf{c}_i$4 the timestamp, and $\mathbf{c}_i$5 the polarity (Xie et al., 2023). A central construction normalizes time across clips,

$\mathbf{c}_i$6

partitions the $\mathbf{c}_i$7 space into spatiotemporal voxels, and assigns each non-empty voxel a motion-sensitive 2D patch

$\mathbf{c}_i$8

followed by a learned feature embedding $\mathbf{c}_i$9 and a 3D coordinate $\mathcal{V}=\{(\mathbf{f}_i,\mathbf{c}_i)\}$0. The resulting representation is the sparse event voxel set

$\mathcal{V}=\{(\mathbf{f}_i,\mathbf{c}_i)\}$1

The importance of this construction is twofold. First, it retains sparsity: complexity scales with the number of informative voxels $\mathcal{V}=\{(\mathbf{f}_i,\mathbf{c}_i)\}$2 rather than with $\mathcal{V}=\{(\mathbf{f}_i,\mathbf{c}_i)\}$3. Second, it aggregates many raw events into a richer local descriptor, which is more robust than pointwise event processing and less redundant than dense event frames or volumes. The same paper explicitly characterizes the voxel set as “a sparse, structured 3D point cloud of macro-events” (Xie et al., 2023).

Two architectures illustrate this line of work. The Event Voxel Set Transformer (EVSTr) processes one voxel set at a time with an Event Voxel Transformer Encoder composed of three Multi-Scale Neighbor Embedding Layers (MNEL), two Voxel Self-Attention Layers (VSAL), and, for long sequences, a segment modeling strategy S$\mathcal{V}=\{(\mathbf{f}_i,\mathbf{c}_i)\}$4TM (Xie et al., 2023). MNEL builds a k-NN graph in $\mathcal{V}=\{(\mathbf{f}_i,\mathbf{c}_i)\}$5-space, uses relative positional vectors

$\mathcal{V}=\{(\mathbf{f}_i,\mathbf{c}_i)\}$6

and performs multi-scale attentive aggregation over nested neighborhood subspaces. VSAL then adds both absolute and relative positional information to self-attention, with pairwise bias

$\mathcal{V}=\{(\mathbf{f}_i,\mathbf{c}_i)\}$7

For action recognition, S$\mathcal{V}=\{(\mathbf{f}_i,\mathbf{c}_i)\}$8TM converts each segment’s encoder output into a token by concatenating max-pooling and average-pooling, prepends a learnable classification token, adds temporal positional embeddings, and applies a shallow transformer encoder.

The earlier EV-VGCNN uses a closely related voxel-wise representation but places it in a graph-CNN pipeline (Deng et al., 2021). Time is normalized into a bounded temporal axis by

$\mathcal{V}=\{(\mathbf{f}_i,\mathbf{c}_i)\}$9

with $\mathcal{U}_i$0, then the $\mathcal{U}_i$1–$\mathcal{U}_i$2–$\mathcal{U}_i$3 cube is divided into voxels of size $\mathcal{U}_i$4. The model retains only the top $\mathcal{U}_i$5 non-empty voxels by event count, treats them as graph vertices with coordinates $\mathcal{U}_i$6, and constructs a 2D patch feature

$\mathcal{U}_i$7

Its Multi-Scale Feature Relational Layer (MFRL) uses separate adjacent and distant k-NN branches to capture local spatial structure and larger motion context.

These event-camera works use “evoxels” to denote a midpoint between dense event frames and raw event points. EVSTr reports state-of-the-art performance with 0.93M parameters and 0.34G MACs on N-Caltech101, while EV-VGCNN reports merely 0.84M parameters (Xie et al., 2023, Deng et al., 2021). This suggests that the evoxel idea is especially valuable when the raw signal is sparse, asynchronous, and noisy, yet the downstream task benefits from locally aggregated motion structure.

3. Environment-centric and adaptive voxel enrichments

In robotics and scene understanding, evoxel-like designs appear as voxels augmented with evidence and traversability attributes rather than only geometry. G-VOM is described as an “enhanced voxel” mapping system for off-road navigation, maintaining a local, ego-centric 3D voxel map with standard configuration $\mathcal{U}_i$8 at $\mathcal{U}_i$9 resolution (Overbye et al., 2021). Its data structure combines a dense 3D lookup table $N_h$0 with a 1D sparse data array $N_h$1, so that occupied voxels store an index into $N_h$2, while free voxels encode miss counts as

$N_h$3

Each occupied voxel record stores

$N_h$4

namely hit counts, miss counts, and lowest observed height.

This augmentation supports a layered interpretation of terrain. Positive obstacles are extracted above the lowest drivable surface in each $N_h$5 column and are classified as hard or soft according to a weighted density estimate. Negative obstacles are inferred around unobserved regions by searching in a cone of directions and testing whether the maximum height discrepancy

$N_h$6

exceeds a threshold. Slope comes from least-squares plane fitting,

$N_h$7

and roughness is the average squared plane-fit residual. The system runs at 9–12 Hz, was validated at 4.5 m/s in autonomous operation and 12 m/s in manual operation, and publishes layered 2D maps such as /height_map, /hard_obstacle_map, /negative_obstacle_map, /slope_map, and /roughness_map (Overbye et al., 2021).

A different enrichment strategy appears in Reconfigurable Voxels for LiDAR. Here the voxel grid remains regular, but each voxel’s neighborhood is adaptively reconfigured by a biased random walk over non-empty neighboring voxels (Wang et al., 2020). The walk probability, number of steps, and transition probabilities are

$N_h$8

Sparse voxels therefore move more and are biased toward denser neighbors. The encoded feature is

$N_h$9

where $N_m$0 is instantiated differently in SECOND and PointPillars. The method is explicitly motivated as a way to improve the stability of voxel features in sparse regions and reports improved detection for small and distant objects without noticeable overhead (Wang et al., 2020).

At still larger scales, Aokana uses chunked SVDAGs, LOD, and streaming to render open-world voxel scenes containing tens of billions of voxels (Fang et al., 4 May 2025). Each chunk stores a $N_m$1 voxel region compressed into its own SVDAG; the three deepest levels are flattened into a $N_m$2 leaf represented by a 64-bit bitmap. Geometry and color are decoupled, LOD is built by aggregating eight child chunks into one parent chunk, and chunk refinement is controlled by

$N_m$3

The rendering pipeline uses compute-shader chunk selection, tile selection with Hi-Z occlusion culling, ray marching, a 64-bit visibility buffer, and a final color-resolve pass. The framework reports up to ninefold memory reduction and rendering speeds up to 4.8 times faster than previous state-of-the-art approaches (Fang et al., 4 May 2025).

Taken together, these systems show that in spatial computing the evoxel idea often means voxel cells endowed with operational state: traversability evidence, adaptive neighborhood structure, or hierarchical rendering metadata. The voxel is no longer merely a sample of occupancy; it becomes an active unit in planning, detection, or visibility computation.

4. Materials-science evoxels and differentiable voxel physics

In materials science, evoxels is also the name of a differentiable-physics framework for voxelized microstructure simulation (Daubner et al., 29 Jul 2025). The framework is Python-based and built around two abstractions: VoxelFields, a NumPy-like container for multiple 3D fields on a shared regular grid, and VoxelGrid, which couples those fields to a backend such as PyTorch or JAX, with boundary conditions, finite-difference stencils, and FFT implementations. A microstructure can therefore be represented directly as voxel-wise fields such as phase indicators $N_m$4, concentrations $N_m$5, indicator functions $N_m$6, diffusivities $N_m$7, mobilities $N_m$8, or multiphase fields $N_m$9.

The framework is designed to operate directly on segmented 3D microscopy data from FIB-SEM, X-ray CT, or tomography, and to solve PDEs on large regular grids of up to $z_{\min}$0 voxels on data-center GPUs (Daubner et al., 29 Jul 2025). It targets effective-property computations, reaction–diffusion systems, and phase-field evolution, including equations such as

$z_{\min}$1

the Gray–Scott system, Cahn–Hilliard,

$z_{\min}$2

and Allen–Cahn dynamics. A key algorithmic choice is semi-implicit Fourier pseudo-spectral timestepping, which leads to updates of the form

$z_{\min}$3

For complex geometries, the framework uses the smoothed boundary method, with a diffuse indicator $z_{\min}$4 that encodes the domain directly on the voxel grid.

What makes these voxel fields “evoxels” in a stronger sense is their differentiability. All FFTs, finite differences, and timestepping updates are assembled from PyTorch or JAX tensor operations, so gradients of scalar objectives with respect to initial fields, parameters, or boundary conditions can be obtained automatically. The framework reports that, for a stiff 3D Cahn–Hilliard problem, its native pseudo-spectral IMEX solver is 1–2 orders of magnitude faster than general-purpose ODE integrators for 1000 time steps, and that it enables $z_{\min}$5 voxels on a 4 GB RTX 500 laptop GPU and $z_{\min}$6 voxels on an NVIDIA RTX A6000 (Daubner et al., 29 Jul 2025).

A closely related use of voxel enrichment appears in VoMP, which predicts volumetric fields of Young’s modulus $z_{\min}$7, Poisson’s ratio $z_{\min}$8, and density $z_{\min}$9 for 3D objects that can be rendered and voxelized (Dagli et al., 27 Oct 2025). The pipeline aggregates per-voxel multi-view DINOv2 features, feeds them through a Geometry Transformer to predict a 2D latent code per active voxel, and decodes those latents through a frozen MatVAE trained on a Material Triplet Dataset of approximately 100,000 real-world $r_k$0 triplets. The decoding constraint is central: each voxel’s material code lies on a learned manifold of physically plausible materials rather than being an unconstrained regression output.

The model is representation-agnostic across meshes, SDFs, NeRFs, and Gaussian splats, supervises the Geometry Transformer through decoded normalized material triplets, and exports per-voxel property fields that can be queried by FEM, Simplicits, XPBD, or MPM (Dagli et al., 27 Oct 2025). On the GVM benchmark it reports an ALDE of 0.38 for Young’s modulus, an ARE of 0.082 for Poisson’s ratio, and an ARE of 0.092 for density, with 3.59 s end-to-end runtime per object. In this setting, evoxels are not just geometric cells; they are simulation-ready material samples.

5. Graph-theoretic and semantic evoxels

A different lineage uses evoxels to encode topology or design semantics. In the atomistic free-volume methodology PANDA, the simulation cell is voxelized into cubic cells of edge length $r_k$1, and an infinitesimal sphere is grown at each voxel center until it either reaches a maximum radius, touches an atomic exclusion sphere, or fails the neighborhood cutoff test (Chapman et al., 2021). The stopping conditions are: $r_k$2; contact with a species-dependent radius $r_k$3; or absence of atoms within $r_k$4, set in that work to $r_k$5. The resulting radius $r_k$6 defines a local free-volume measure

$r_k$7

These sphere-augmented voxels are then converted into a graph. Each voxel becomes a node if its sphere overlaps with a neighbor’s sphere, with edges defined by

$r_k$8

Connected components correspond to free-volume elements such as isolated pockets, channels, or surface trenches. To quantify topology, PANDA defines the SGOP-V order parameter, which combines local free volumes with the degree distribution of each subgraph. The method is validated on crystal phases, vacancy defects, epitaxial surface growth, vacancy diffusion, and million-atom porous Al simulations, and reports runtime of less than 30 seconds for systems larger than 1 million atoms using 36 threads on an Intel Xeon E5-2695 v4 (Chapman et al., 2021). Here the evoxel is a sampling point in free space endowed with both geometric scale and graph connectivity.

In CAD/CAE integration, XVoxels—eXtended Voxels—hybridize feature models and voxel models into semantic voxels (Li et al., 2023). An XVoxel model contains a feature list plus a voxel attribute array in which each voxel stores a linked list of tuples

$r_k$9

Feature nature denotes additive or subtractive Boolean action, and occupancy completeness denotes full or partial coverage. Since the full feature history is retained, a voxel’s effective state can be recovered from the suffix of the feature list that remains relevant after the last fully occupied feature.

This design creates a direct mapping between design models and analysis models. When a feature parameter changes, the method deletes and re-voxelizes only the affected feature, marks changed cells as active voxels, and recomputes stiffness matrices only for those voxels. Analysis is performed with the Finite Cell Method (FCM) on a fixed background voxel grid, using a characteristic function

$$(\text{feature index},\text{feature nature},\text{occupancy completeness})$$0

and a global system

$$(\text{feature index},\text{feature nature},\text{occupancy completeness})$$1

For optimization, the work derives feature-parameter sensitivities of stiffness and compliance as boundary integrals over feature boundaries, enabling gradient-based updates via GCMMA. Across case studies, the reported computational efficiency improvement reaches up to 55.8 times the existing FCM method (Li et al., 2023).

PANDA and XVoxels differ sharply in application, but both exemplify a defining evoxel move: the voxel is upgraded from a spatial sample to a semantic carrier. In PANDA, semantics are topological and defect-centered; in XVoxels, they are feature-historical and CAD-aware.

6. Recurrent principles, limitations, and directions

Several recurring principles cut across these disparate formulations. First, evoxel systems usually preserve the algorithmic convenience of voxelization while compensating for its naïve limitations by attaching higher-order state. Event-camera methods add motion-sensitive local features and positional encodings (Xie et al., 2023, Deng et al., 2021). Robotics systems add occupancy evidence, surface height, or adaptive neighborhood structure (Overbye et al., 2021, Wang et al., 2020). Materials frameworks store fields, constitutive parameters, or latent material codes (Daubner et al., 29 Jul 2025, Dagli et al., 27 Oct 2025). CAD/CAE systems store feature provenance and Boolean history (Li et al., 2023). This suggests that evoxels are less a single representation than a design pattern: voxelization plus task-aware augmentation.

Second, many evoxel systems are explicitly motivated by the trade-off between dense regularity and sparse reality. Event-camera work argues that dense event frames remove sparsity and force CNN-like computation over every pixel, while pure point clouds can be too large and too noise-sensitive (Xie et al., 2023). G-VOM contrasts its $$(\text{feature index},\text{feature nature},\text{occupancy completeness})$$2 lookup-table access with octrees’ pointer-heavy structure and lack of rich payload for empty space (Overbye et al., 2021). Reconfigurable Voxels keep the regular grid but deform the neighborhood used for feature aggregation (Wang et al., 2020). Aokana keeps chunk-local regularity while using SVDAG compression, LOD, and streaming to avoid full-scene residency in VRAM (Fang et al., 4 May 2025).

Third, the literature makes clear that evoxels are not free of limitations. Event-graph methods still assume that the input stream is sufficiently object-relevant for Euclidean k-NN neighborhoods to be meaningful, and they remain synchronous rather than fully asynchronous (Deng et al., 2021). G-VOM has no explicit long-term memory and is intentionally local and rolling (Overbye et al., 2021). The differentiable-physics framework evoxels is restricted to uniform voxel grids, focuses on diffusion/reaction–diffusion/phase-field PDEs, and notes that time-varying Dirichlet conditions are more challenging in its FFT-based schemes (Daubner et al., 29 Jul 2025). VoMP assumes isotropic per-voxel materials and relies on part-level supervision (Dagli et al., 27 Oct 2025). XVoxels remain grid-based, depend on the initial parametric feature model, and do not yet address dimensional reduction to shells or beams (Li et al., 2023). Aokana is static in its current form and does not provide runtime voxel modification or transparent voxels (Fang et al., 4 May 2025).

A further misconception is that “evoxels” must be dense cubic grids. The surveyed literature includes sparse voxel sets of tokens, voxel-centered graphs, regular grids with sparse data arrays, chunked SVDAG hierarchies, and voxel fields embedded in autograd pipelines. Another misconception is that the concept is tied to a single discipline. The term spans neuromorphic vision, robotics, materials informatics, atomistic analysis, CAD/CAE integration, and large-scale rendering, with each field adapting the voxel abstraction to its own operator class—transformers, graph convolutions, planners, PDE solvers, or renderers.

Future directions are likewise domain-specific. Event-camera work points toward multi-camera or depth-aware neuromorphic sensing, SLAM, tracking, optical flow, and multimodal fusion (Xie et al., 2023). The differentiable-physics framework points toward coupled chemo-mechanics, electrochemistry, stronger inverse-design workflows, and multi-scale coupling (Daubner et al., 29 Jul 2025). VoMP points toward anisotropic and nonlinear constitutive behavior and tighter coupling to differentiable simulators (Dagli et al., 27 Oct 2025). These trajectories suggest that evoxels are becoming increasingly important where discrete spatial structure must remain compatible with semantics, optimization, or learned representations.

In that broad sense, evoxels are best understood as enriched volumetric primitives: cells of a spatial discretization whose payload is sufficiently expressive to support reasoning, inference, simulation, or control beyond what a conventional occupancy voxel can provide.