Part-Aware Spatial Configuration Solver

• Part-Aware Spatial Configuration Solvers decompose objects or scenes into constituent parts and reason over their local geometries and relationships.
• They integrate explicit constraint grounding with implicit neural inference to assemble, refine, or match spatial arrangements.
• Their applications span 2D/3D recognition, pose estimation, and scene assembly, demonstrating robust performance improvements in spatial tasks.

A Part-Aware Spatial Configuration Solver denotes, in the cited literature, a mechanism that infers, assembles, refines, or matches spatial arrangements by decomposing an object, body, or scene into parts and reasoning over their local geometry, mutual relations, and composition rather than relying only on holistic representations. The concept appears in multiple technical forms: as an efficient solver for Localized Deformable Part Models in 2D recognition (Zhang et al., 2018), as a two-stage LiDAR detector that predicts intra-object part locations and refines proposals from their arrangement (Shi et al., 2019), as dual-level body/part prototype reasoning for one-shot skeleton action recognition (Chen et al., 2022), as sparse part-level query refinement for multi-person pose regression (Xiao et al., 2022), as a procedural engine that converts part-level scene relations into geometric constraints and collision-free 3D layouts (Bai et al., 8 Mar 2026), and as identity-aligned or layout-conditioned generation systems that solve which semantic part occupies which slot and where that part should be placed (Slim et al., 19 May 2026, Hao et al., 10 Jun 2026). Closely related work also supplies perceptual and evaluative substrates for such solvers, notably part-aware 3D grounding (Fei et al., 2024) and action-conditioned spatial repair benchmarks (Liu et al., 29 May 2026).

1. Definition and conceptual scope

The literature does not present a single canonical formulation of the term. Instead, it spans a spectrum from explicit constraint-grounding systems to differentiable amortized inference architectures. In “Deformable Part Networks,” the network is explicitly interpreted as “an efficient solver” for Localized Deformable Part Models, with feedforward computation realizing nested maximization over part placement, localization window, and image position (Zhang et al., 2018). In PARSE, the “Part-Aware Spatial Configuration Solver” is a procedural engine that instantiates a Part-centric Assembly Graph into a 3D scene by topological traversal, support-region pruning, part-level alignment, collision validation, and physics refinement (Bai et al., 8 Mar 2026). By contrast, the skeleton and generation papers emphasize learned, implicit solving: PPGN “does not formulate an explicit symbolic constraint solver over body parts,” but learns an implicit solver through graph-based representation learning, attentional part selection, and prototype matching (Chen et al., 2022), while CompoSE is “primarily implicit learned solving through diffusion” with a small explicit post-generation layout refinement step (Slim et al., 19 May 2026).

A concise typology is useful for situating the main variants.

Work	Domain	Solver interpretation
DPN (Zhang et al., 2018)	2D object recognition	Efficient solver for LDPMs
Part-$A^2$ Net (Shi et al., 2019)	3D object detection from LiDAR	Part-aware stage plus part-aggregation stage
PPGN (Chen et al., 2022)	One-shot skeleton action recognition	Implicit solver via graph learning, attention, and prototype matching
QueryPose (Xiao et al., 2022)	Multi-person pose regression	Sparse structured prediction with part-level queries
PARSE (Bai et al., 8 Mar 2026)	3D indoor scene assembly	Constraint-grounding and incremental scene assembly
CompoSE (Slim et al., 19 May 2026)	3D shape synthesis/editing	Solve detailed part geometry from coarse part layouts
ISAP-3D (Hao et al., 10 Jun 2026)	Part-aware 3D generation	Identity-aligned one-to-one slot modelling

A recurrent misconception is that “part-aware” automatically implies explicit semantic parts and symbolic constraints. The cited work shows otherwise. Part-$A^2$ Net uses “intra-object part locations” that are normalized coordinates inside the object box rather than semantic parts such as wheels or doors (Shi et al., 2019). QueryPose uses learnable part-level queries associated with a chosen body-part partition, but the method remains a neural latent structured predictor rather than an analytic kinematic solver (Xiao et al., 2022). Conversely, PARSE is explicitly procedural and geometric, yet its representation is still grounded in object and part nodes rather than a single monolithic scene graph (Bai et al., 8 Mar 2026).

2. Formal representations of configuration

One major formal lineage begins with latent-variable part models. DPN extends standard DPMs by introducing localization as an additional latent variable $h$ beside the deformable-part configuration $z$, yielding the LDPM objective

$$\min_{\mathbf{w}\in\mathcal{W}} \left\{ \sum_{(x,y)\in\mathcal{X}\times\mathcal{Y}} \ell\left( y,\, \max_{(h,z)\in\mathcal{H}\times\mathcal{Z}} \Big\{\mathbf{w}^T\phi(x,h,z)\Big\} \right) \right\},$$

and a localized deformable matching rule

$$\psi(x,h(i),n) = \max_{j\in h(i)} \Big\{ \alpha(h,j,n)\otimes\psi(x,j,n)\oplus\beta(h,j,n) \Big\}.$$

The resulting network realizes search over part deformation, support window, and image location in a hierarchical composition process (Zhang et al., 2018).

A second formal lineage represents configuration by normalized intra-object coordinates. Part-$A^2$ Net defines, for each foreground LiDAR point inside a box, normalized part coordinates

$$x^{(part)} = \frac{x^{(t)}}{w} + 0.5,\qquad y^{(part)} = \frac{y^{(t)}}{l} + 0.5,\qquad z^{(part)} = \frac{z^{(p)} - z^{(c)}}{h} + 0.5,$$

so that the object center maps to $(0.5,0.5,0.5)$. These predicted intra-object part locations are then pooled inside proposal-aligned voxel grids and aggregated for proposal scoring and refinement (Shi et al., 2019).

In skeleton-based one-shot recognition, PPGN formalizes a meta-task $T$ with support set

$A^2$0

query set

$A^2$1

and a spatio-temporal skeleton graph input

$A^2$2

A body-level encoder produces

$A^2$3

part-level sampling yields $A^2$4, average pooling gives

$A^2$5

attention weights are computed as

$A^2$6

and the fused embedding

$A^2$7

is matched by cosine distance for classification (Chen et al., 2022).

QueryPose represents each person by one instance-level query $A^2$8 and $A^2$9 part-level queries $h$0. At stage $h$1,

$h$2

so configuration is carried by persistent part-query states updated from ROI-conditioned part embeddings and self-attention across parts (Xiao et al., 2022).

In part-aware 3D generation, ISAP-3D formalizes conditions

$h$3

part geometries

$h$4

a coarse voxel prior $h$5, and per-part bounding boxes

$h$6

with the structured mapping

$h$7

Its identity-conditioned layout transformer updates slot $h$8 by

$h$9

which first grounds the slot in its own identity condition, then in other slots, then in global context (Hao et al., 10 Jun 2026).

PARSE differs in that it does not provide a single global optimization objective. The paper explicitly states that the solver works by progressive refinement of feasible pose space: start from a 2D candidate region on the support surface, exclude occupied regions, apply object-level relations, resolve part-level relations into geometric constraints such as parallelism and contact, sample a pose from the remaining subspace, validate for collisions and semantic plausibility, and refine with physics simulation in Sapien (Bai et al., 8 Mar 2026). CompoSE likewise formalizes the task as

$z$0

with VAE-encoded part latents

$z$1

and box controls encoded in the same latent space

$z$2

so that coarse layout primitives serve as direct geometric conditions for part synthesis (Slim et al., 19 May 2026).

3. Core algorithmic patterns

Across otherwise heterogeneous domains, several recurrent mechanisms define the solver family. The first is structured decomposition. PPGN uses a fixed set of $z$3 part graphs generated by semantic, symmetry, and mixture partitions (Chen et al., 2022). QueryPose studies 17-part, 13-part, 7-part, and 5-part decompositions, with the 7-part rigid local-region design performing best (Xiao et al., 2022). ISAP-3D instantiates exactly $z$4 prediction queries for $z$5 semantic parts and ties each slot to one semantic identity condition $z$6 (Hao et al., 10 Jun 2026). PARSE represents scenes with object nodes $z$7, part nodes $z$8, object-level spatial edges, and part-level geometric edges inside a support-driven directed acyclic graph (Bai et al., 8 Mar 2026). CompoSE uses one oriented bounding box per intended part and synthesizes one part per input primitive (Slim et al., 19 May 2026).

The second is local-global coupling. DPN composes local part evidence into higher-level parts by hierarchical LDPM units (Zhang et al., 2018). PPGN first computes globally contextualized body-level features $z$9, then samples local part tensors from that feature map, so local part reasoning is explicitly conditioned on whole-body context (Chen et al., 2022). QueryPose extracts local part embeddings from a person ROI through the Spatial Part Embedding Generation Module and then uses multi-head self-attention across part queries to recover global person structure (Xiao et al., 2022). ISAP-3D preserves identity alignment across semantic, spatial, and geometric stages by combining local identity conditions $$\min_{\mathbf{w}\in\mathcal{W}} \left\{ \sum_{(x,y)\in\mathcal{X}\times\mathcal{Y}} \ell\left( y,\, \max_{(h,z)\in\mathcal{H}\times\mathcal{Z}} \Big\{\mathbf{w}^T\phi(x,h,z)\Big\} \right) \right\},$$0 with global context $$\min_{\mathbf{w}\in\mathcal{W}} \left\{ \sum_{(x,y)\in\mathcal{X}\times\mathcal{Y}} \ell\left( y,\, \max_{(h,z)\in\mathcal{H}\times\mathcal{Z}} \Big\{\mathbf{w}^T\phi(x,h,z)\Big\} \right) \right\},$$1 in both layout prediction and geometry synthesis (Hao et al., 10 Jun 2026). CompoSE alternates local blocks, where one part is processed independently and conditioned on its own box, with global blocks, where all part tokens attend jointly under an enclosing OBB (Slim et al., 19 May 2026).

The third is solver realization. In PARSE, solving is explicit: topological assembly order comes from the PAG DAG, support relations define sequential placement, part references may be specified or inferred, and validity is enforced by geometric constraints, collision checks, semantic plausibility tests, and final physics refinement (Bai et al., 8 Mar 2026). In DPN, solving is also explicit in a different sense: latent maximization over $$\min_{\mathbf{w}\in\mathcal{W}} \left\{ \sum_{(x,y)\in\mathcal{X}\times\mathcal{Y}} \ell\left( y,\, \max_{(h,z)\in\mathcal{H}\times\mathcal{Z}} \Big\{\mathbf{w}^T\phi(x,h,z)\Big\} \right) \right\},$$2, $$\min_{\mathbf{w}\in\mathcal{W}} \left\{ \sum_{(x,y)\in\mathcal{X}\times\mathcal{Y}} \ell\left( y,\, \max_{(h,z)\in\mathcal{H}\times\mathcal{Z}} \Big\{\mathbf{w}^T\phi(x,h,z)\Big\} \right) \right\},$$3, and $$\min_{\mathbf{w}\in\mathcal{W}} \left\{ \sum_{(x,y)\in\mathcal{X}\times\mathcal{Y}} \ell\left( y,\, \max_{(h,z)\in\mathcal{H}\times\mathcal{Z}} \Big\{\mathbf{w}^T\phi(x,h,z)\Big\} \right) \right\},$$4 is compiled into Deformable Maxout, maxout over windows, and top-layer spatial selection (Zhang et al., 2018). By contrast, PPGN describes an “implicit solver for part-aware spatial configuration” through graph-based representation learning and attentional part fusion (Chen et al., 2022), while CompoSE and QueryPose rely on learned latent compatibility and iterative refinement rather than closed-form geometric objectives (Slim et al., 19 May 2026, Xiao et al., 2022). This suggests that “solver” in the literature covers both constraint satisfaction and amortized neural inference.

A fourth recurrent mechanism is selection or gating over candidate part explanations. PPGN computes sample-dependent, class-agnostic part importance weights $$\min_{\mathbf{w}\in\mathcal{W}} \left\{ \sum_{(x,y)\in\mathcal{X}\times\mathcal{Y}} \ell\left( y,\, \max_{(h,z)\in\mathcal{H}\times\mathcal{Z}} \Big\{\mathbf{w}^T\phi(x,h,z)\Big\} \right) \right\},$$5 and rescales each part descriptor before fusion (Chen et al., 2022). QueryPose uses the Selective Iteration Module: $$\min_{\mathbf{w}\in\mathcal{W}} \left\{ \sum_{(x,y)\in\mathcal{X}\times\mathcal{Y}} \ell\left( y,\, \max_{(h,z)\in\mathcal{H}\times\mathcal{Z}} \Big\{\mathbf{w}^T\phi(x,h,z)\Big\} \right) \right\},$$6 thereby retaining prior latent state when current stage evidence is noisy (Xiao et al., 2022). ISAP-3D removes slot interchangeability by construction through one semantic identity $$\min_{\mathbf{w}\in\mathcal{W}} \left\{ \sum_{(x,y)\in\mathcal{X}\times\mathcal{Y}} \ell\left( y,\, \max_{(h,z)\in\mathcal{H}\times\mathcal{Z}} \Big\{\mathbf{w}^T\phi(x,h,z)\Big\} \right) \right\},$$7 one dedicated slot, which is an architectural alternative to explicit matching or permutation-invariant objectives (Hao et al., 10 Jun 2026).

4. Major application domains

In recognition and localization, part-aware spatial configuration solving appears as a method for robustly matching structured evidence under deformation, sparsity, or low-data transfer. DPN solves 2D recognition by searching over object locations, scales, and deformations of constituent parts and composing the best-scoring arrangement into a pose-invariant representation (Zhang et al., 2018). Part-$$\min_{\mathbf{w}\in\mathcal{W}} \left\{ \sum_{(x,y)\in\mathcal{X}\times\mathcal{Y}} \ell\left( y,\, \max_{(h,z)\in\mathcal{H}\times\mathcal{Z}} \Big\{\mathbf{w}^T\phi(x,h,z)\Big\} \right) \right\},$$8 Net detects 3D objects from LiDAR by predicting part coordinates for foreground points, then evaluating whether the arrangement of those part coordinates is compatible with a valid proposal (Shi et al., 2019). PPGN targets one-shot skeleton-based action recognition, where subtle distinctions among novel classes may depend on hands or arms rather than only body-level motion, so local part motifs must transfer across classes (Chen et al., 2022). QueryPose addresses sparse end-to-end multi-person pose regression by representing each person with several spatial-aware part-level queries instead of a single monolithic instance token (Xiao et al., 2022).

In 3D generation and editing, the solver view centers on correspondence between semantic identity, layout, and geometry. ISAP-3D argues that unstable part allocation arises from “identity-slot permutation freedom,” so part-aware generation requires identity-aligned one-to-one slot modelling, identity-conditioned layout prediction, and layout-conditioned geometry synthesis (Hao et al., 10 Jun 2026). CompoSE treats the user’s set of oriented bounding boxes as the primary structured specification of object composition and solves for detailed part geometry conditioned on box layout and a global text prompt, with localized editing implemented through latent freezing and identity-preserving resizing through key/value reinjection (Slim et al., 19 May 2026).

In scene assembly, PARSE is the clearest explicit part-aware solver. It turns a Part-centric Assembly Graph into a 3D scene by processing object nodes in topological order, grounding part-level relations such as on, in, against, and aligned with into support-plane, coplanarity, contact, and enclosure checks, then validating collisions and extracting a part-level contact graph after simulation (Bai et al., 8 Mar 2026). The paper emphasizes that object-level scene graphs and linguistic prepositions are too coarse because they do not specify which regions actually support, contain, or contact one another (Bai et al., 8 Mar 2026).

Perceptual grounding and benchmark work supplies adjacent components. Kestrel is “not itself a full Part-Aware Spatial Configuration Solver,” but it provides a language-conditioned part localization module over 3D point clouds, with [SEG]-conditioned mask decoding and two tasks—Part-Aware Point Grounding and Part-Aware Point Grounded Captioning—that expose part and material entities a downstream solver could reason over (Fei et al., 2024). SpatialAct is likewise “not a part-aware benchmark in the strict sense,” but it evaluates whether models can translate spatial reasoning into move, rotate, and scale actions across iterative simulator feedback, which is directly relevant to stateful configuration repair (Liu et al., 29 May 2026).

5. Empirical evidence

The empirical record shows that part-aware formulations often outperform holistic or weakly structured baselines when the target problem depends on local geometry or stable correspondence. On affNIST, the abstract reports that a 17-layer DPN outperforms CapsNets by $$\min_{\mathbf{w}\in\mathcal{W}} \left\{ \sum_{(x,y)\in\mathcal{X}\times\mathcal{Y}} \ell\left( y,\, \max_{(h,z)\in\mathcal{H}\times\mathcal{Z}} \Big\{\mathbf{w}^T\phi(x,h,z)\Big\} \right) \right\},$$9 and STNs by $$\psi(x,h(i),n) = \max_{j\in h(i)} \Big\{ \alpha(h,j,n)\otimes\psi(x,j,n)\oplus\beta(h,j,n) \Big\}.$$0, and with only 3 images per class for training it achieves $$\psi(x,h(i),n) = \max_{j\in h(i)} \Big\{ \alpha(h,j,n)\otimes\psi(x,j,n)\oplus\beta(h,j,n) \Big\}.$$1, a $$\psi(x,h(i),n) = \max_{j\in h(i)} \Big\{ \alpha(h,j,n)\otimes\psi(x,j,n)\oplus\beta(h,j,n) \Big\}.$$2 improvement over the second best method (Zhang et al., 2018). In LiDAR detection, Part-$$\psi(x,h(i),n) = \max_{j\in h(i)} \Big\{ \alpha(h,j,n)\otimes\psi(x,j,n)\oplus\beta(h,j,n) \Big\}.$$3 Net reports that adding part prediction improves stage-II refinement: on KITTI val, car, IoU $$\psi(x,h(i),n) = \max_{j\in h(i)} \Big\{ \alpha(h,j,n)\otimes\psi(x,j,n)\oplus\beta(h,j,n) \Big\}.$$4, stage I + II without part prediction reaches recall $$\psi(x,h(i),n) = \max_{j\in h(i)} \Big\{ \alpha(h,j,n)\otimes\psi(x,j,n)\oplus\beta(h,j,n) \Big\}.$$5 and moderate AP $$\psi(x,h(i),n) = \max_{j\in h(i)} \Big\{ \alpha(h,j,n)\otimes\psi(x,j,n)\oplus\beta(h,j,n) \Big\}.$$6, while stage I + II with part prediction reaches recall $$\psi(x,h(i),n) = \max_{j\in h(i)} \Big\{ \alpha(h,j,n)\otimes\psi(x,j,n)\oplus\beta(h,j,n) \Big\}.$$7 and moderate AP $$\psi(x,h(i),n) = \max_{j\in h(i)} \Big\{ \alpha(h,j,n)\otimes\psi(x,j,n)\oplus\beta(h,j,n) \Big\}.$$8; the mean part-location error for cars is $$\psi(x,h(i),n) = \max_{j\in h(i)} \Big\{ \alpha(h,j,n)\otimes\psi(x,j,n)\oplus\beta(h,j,n) \Big\}.$$9, and the Pearson correlation between part error and box error is $A^2$0 overall (Shi et al., 2019).

In low-data action recognition, PPGN reaches $A^2$1 on NTU RGB+D 120 with 100 training classes and $A^2$2 on NW-UCLA, outperforming several ProtoNet variants and prior metric-learning baselines (Chen et al., 2022). Its ablations isolate the solver ingredients: replacing the last five layers of the ST-GCN baseline with part-level modeling raises performance from $A^2$3 to $A^2$4, and adding attentional part fusion reaches $A^2$5; varying the number of part graphs shows $A^2$6 performs best at $A^2$7, whereas $A^2$8 drops to $A^2$9 (Chen et al., 2022). In sparse pose estimation, QueryPose reports $$x^{(part)} = \frac{x^{(t)}}{w} + 0.5,\qquad y^{(part)} = \frac{y^{(t)}}{l} + 0.5,\qquad z^{(part)} = \frac{z^{(p)} - z^{(c)}}{h} + 0.5,$$0 AP on MS COCO mini-val and $$x^{(part)} = \frac{x^{(t)}}{w} + 0.5,\qquad y^{(part)} = \frac{y^{(t)}}{l} + 0.5,\qquad z^{(part)} = \frac{z^{(p)} - z^{(c)}}{h} + 0.5,$$1 AP on CrowdPose test, while the key ablations show SPEGM improves AP from $$x^{(part)} = \frac{x^{(t)}}{w} + 0.5,\qquad y^{(part)} = \frac{y^{(t)}}{l} + 0.5,\qquad z^{(part)} = \frac{z^{(p)} - z^{(c)}}{h} + 0.5,$$2 to $$x^{(part)} = \frac{x^{(t)}}{w} + 0.5,\qquad y^{(part)} = \frac{y^{(t)}}{l} + 0.5,\qquad z^{(part)} = \frac{z^{(p)} - z^{(c)}}{h} + 0.5,$$3 and SIM further improves it to $$x^{(part)} = \frac{x^{(t)}}{w} + 0.5,\qquad y^{(part)} = \frac{y^{(t)}}{l} + 0.5,\qquad z^{(part)} = \frac{z^{(p)} - z^{(c)}}{h} + 0.5,$$4 (Xiao et al., 2022).

In 3D generation, ISAP-3D reports CD $$x^{(part)} = \frac{x^{(t)}}{w} + 0.5,\qquad y^{(part)} = \frac{y^{(t)}}{l} + 0.5,\qquad z^{(part)} = \frac{z^{(p)} - z^{(c)}}{h} + 0.5,$$5, F $$x^{(part)} = \frac{x^{(t)}}{w} + 0.5,\qquad y^{(part)} = \frac{y^{(t)}}{l} + 0.5,\qquad z^{(part)} = \frac{z^{(p)} - z^{(c)}}{h} + 0.5,$$6, Part-IoU $$x^{(part)} = \frac{x^{(t)}}{w} + 0.5,\qquad y^{(part)} = \frac{y^{(t)}}{l} + 0.5,\qquad z^{(part)} = \frac{z^{(p)} - z^{(c)}}{h} + 0.5,$$7, and NMI $$x^{(part)} = \frac{x^{(t)}}{w} + 0.5,\qquad y^{(part)} = \frac{y^{(t)}}{l} + 0.5,\qquad z^{(part)} = \frac{z^{(p)} - z^{(c)}}{h} + 0.5,$$8, compared with OmniPart at CD $$x^{(part)} = \frac{x^{(t)}}{w} + 0.5,\qquad y^{(part)} = \frac{y^{(t)}}{l} + 0.5,\qquad z^{(part)} = \frac{z^{(p)} - z^{(c)}}{h} + 0.5,$$9, F $(0.5,0.5,0.5)$0, Part-IoU $(0.5,0.5,0.5)$1, and NMI $(0.5,0.5,0.5)$2. The paper interprets the NMI gain as the clearest evidence that identity-slot alignment improves part assignment stability, while lower Part-IoU indicates better part separation and less merging (Hao et al., 10 Jun 2026). CompoSE reports, for the 2048-token model, Object-FID $(0.5,0.5,0.5)$3, Part-FID $(0.5,0.5,0.5)$4, Object-IoU $(0.5,0.5,0.5)$5, Voxel-IoU $(0.5,0.5,0.5)$6, Part-IoU $(0.5,0.5,0.5)$7, and CLIP-Score $(0.5,0.5,0.5)$8, and its layout-encoding ablation shows that shared SDF/VAE latent encoding yields Part-IoU $(0.5,0.5,0.5)$9, far above parametric encoding at $T$0 and DGCNN point encoding at $T$1 (Slim et al., 19 May 2026).

In scene assembly and grounded supervision, PARSE uses its solver to construct PARSE-10K, a dataset of 10,000 indoor scenes with average $T$2 objects per scene and dense part-level contact structure (Bai et al., 8 Mar 2026). Fine-tuning Qwen3-VL on PARSE-10K yields $T$3 on Visual Relation MCQ and $T$4 on Part-level Contact MCQ, and the paper also reports user-study preferences for PAG-conditioned generation in Complexity $T$5, Realism $T$6, and Contact Fidelity $T$7 (Bai et al., 8 Mar 2026). Kestrel provides the strongest evidence on the perception side: on 3DCoMPaT-GRIN Vanilla test, Kestrel-7B reaches Part $T$8, Material $T$9, Composition $A^2$00, and Overall $A^2$01 IoU, compared with PointLLM-SG-7B at Overall $A^2$02 (Fei et al., 2024). SpatialAct, finally, demonstrates the limits of current general-purpose VLMs as iterative configuration solvers: the strongest VLM only achieves $A^2$03 Repair Rate and $A^2$04 Scene Success Rate in multi-turn simulator-grounded refinement, whereas human participants achieve $A^2$05 Repair Rate and $A^2$06 Scene Success Rate (Liu et al., 29 May 2026).

6. Limitations, controversies, and open directions

A central controversy concerns what counts as “solving.” PARSE is explicit that its solver is a hybrid constraint-satisfaction and sampling procedure rather than a global energy minimizer, and the paper does not provide an explicit conflict-resolution mechanism or relaxation schedule for contradictory PAG specifications (Bai et al., 8 Mar 2026). PPGN, CompoSE, and QueryPose instead embed configuration reasoning in learned representations and attention, so they improve matching or generation without explicitly encoding hard relations such as “left hand must be near torso while right hand moves upward” or CAD-grade contact constraints (Chen et al., 2022, Slim et al., 19 May 2026, Xiao et al., 2022). Taken together, these works suggest that the field uses “solver” for both declarative constraint grounding and amortized neural inference, and that the distinction matters when guarantees, interpretability, or physical validity are required.

A second limitation is the dependence on predefined structure. PPGN’s part decomposition is fixed and heuristic rather than learned (Chen et al., 2022). QueryPose’s latent parts depend on a hand-chosen partition scheme, and the exact mapping from part queries to all keypoints is not fully formalized in the paper text (Xiao et al., 2022). Part-$A^2$07 Net’s “parts” are normalized coordinates inside a box rather than semantic components, which is powerful for geometry but not equivalent to semantic part reasoning (Shi et al., 2019). ISAP-3D relies on a curated closed semantic vocabulary and about 8K curated objects with around 70 unified semantic part categories, so its identity-slot alignment is learned under controlled semantics rather than open-vocabulary composition (Hao et al., 10 Jun 2026).

A third limitation is that many current systems are only partially physical or relational. CompoSE explicitly does not expose a symbolic or optimization-based solver for contact, collision avoidance, exact pairwise relations, support, or physical stability, and it is capped at 8 parts per shape due to global attention memory cost (Slim et al., 19 May 2026). PARSE acknowledges that its directional-face contact model handles near-parallel surfaces well but may deviate on highly oblique contacts, and containment for in is validated by multi-directional ray-casting rather than an exact volumetric solver (Bai et al., 8 Mar 2026). SpatialAct shows that even strong VLMs that perform well on isolated spatial tasks still lack robust spatial state tracking under action-induced environment changes (Liu et al., 29 May 2026).

A final limitation is perceptual incompleteness. Kestrel is a strong part-aware grounding substrate, but it does not predict part graphs, support relations, attachment structure, symmetry, or articulation, and it focuses on single objects rather than multi-object interaction (Fei et al., 2024). SpatialAct is likewise object-level rather than part-aware, with move, rotate, and scale actions over whole objects and no explicit articulation or contact compatibility (Liu et al., 29 May 2026). A plausible implication is that a full part-aware spatial configuration solver will require at least three layers simultaneously: grounded part perception, explicit or implicit relational reasoning over those parts, and a stateful repair or generation loop that updates beliefs after actions or sampling. The cited literature provides strong instances of each layer separately, but only a few systems—most notably PARSE—integrate them into a single procedural realization (Bai et al., 8 Mar 2026).