SWAP Framework in Structure-Aware Planning

Updated 21 March 2026

SWAP Framework is a structure-aware planning paradigm that integrates geometric, semantic, and symbolic representations to enhance robustness and interpretability.
It leverages techniques like scene graphs, BEV maps, and graph-based neural attention to optimize trajectory synthesis and reduce computational complexity.
The framework demonstrates significant performance gains in domains such as autonomous driving, inspection, and robotics by enforcing domain-specific structural constraints.

Structure-aware planning refers to a family of frameworks and algorithms in which explicit structural information – geometric, topological, semantic, or relational – is incorporated into the planning process to achieve greater robustness, interpretability, sample efficiency, and alignment with domain constraints. This paradigm spans domains such as autonomous driving, semantic inspection, structured assembly, task and motion planning, sequence modeling, and symbolic reasoning. The key innovation is the fusion of structure representation (e.g., scene graphs, skeletons, symbolic graphs, semantic priors, temporal logic) with the core planning process, resulting in plans or policies that are interpretable, scalable, and better able to anticipate or adapt to the complexities of real-world environments.

1. Formal Representations of Structure in Planning

Structure-aware planning relies on defining explicit representations that capture the organizational, geometric, or semantic characteristics of the environment or task. Examples include:

Semantic and scene graphs: Used in predictive inspection and situationally-aware navigation, these encode the spatial/topological relationships between objects, rooms, and functional entities (Dharmadhikari et al., 6 Jun 2025, Ejaz et al., 8 Aug 2025).
BEV (Bird's-Eye View) maps: For autonomous driving, dense pixel-level maps encode both geometric layout and class structure, incorporating uncertainty measures at each location (Ryu et al., 28 Nov 2025).
Skeletons and medial axes: In motion planning, environment geometry is reduced to connectivity graphs or skeletons for efficient initialization and path search (Ryu, 28 May 2025).
Procrustean graphs (p-graphs): Abstract language-theoretic representations that generalize plans, filters, and hybrid automata, supporting unified reasoning about alternating actions and observations (Saberifar et al., 2018).
Entailment graphs and symbolic state arrays: Used in symbolic task planning and language-model-based reasoning, these structures facilitate verifiable, dependency-aware planning (Xiong et al., 2024, Chen et al., 2 Jan 2026).

These representations serve as the substrate for both reasoning (symbolic or geometric) and for constraining the planning or learning algorithms to comply with environment- or task-specific structure.

2. Structural Fusion in Modern Planning Pipelines

State-of-the-art approaches in structure-aware planning go beyond mere structure representation, embedding structural cues deeply into the planning pipeline:

Uncertainty-aware planning: SUPER-AD injects pixelwise aleatoric uncertainty estimates (means and standard deviations over BEV logits) into a trajectory-scoring process, resulting in safety weights that penalize unsafe or ambiguous regions. Lane structure is regularized via intent and centerline-adherence losses, enforcing compliance with traffic norms while permitting flexible maneuvers (Ryu et al., 28 Nov 2025).
Pattern discovery and prediction: In Semantics-aware Predictive Planning (SPP), repeating subgraph patterns are identified using MDL-based extensions of SUBDUE. This enables online prediction of unobserved, yet plausible, structures in partially explored environments, which are then explicitly targeted in subsequent planning steps (Dharmadhikari et al., 6 Jun 2025).
Hierarchical and constraint-driven assignment: Symbolic assembly or reasoning tasks use state representations that explicitly encode component dependencies, supporting minimal-change reassignment, frontier computation, and domain rule validation to optimize both plan stability and workload balancing in dynamic, collaborative settings (Chen et al., 2 Jan 2026, Xiong et al., 2024).
Task and motion planning under logic constraints: Structured-MoE STL planners parse Signal Temporal Logic (STL) specifications into temporally anchored embeddings, routing autoregressive planning queries through Mixture-of-Experts subnetworks specialized for STL operators and time bands. This delivers horizon- and compositionality-aware trajectory synthesis with explicit logic verification (Ye et al., 16 Sep 2025).
Graph-based neural attention: In manipulation planning, GAIDE encodes both robot embodiment and workspace geometry as a single graph, using learned attention masks to restrict transformer attention to spatially and kinematically feasible neighborhoods, improving sample efficiency and solution quality (Soleymanzadeh et al., 3 Mar 2026).

3. Methods for Incorporating Structure into Planning Algorithms

Methods for integrating structure into the planning process include:

Structure-aware initialization and heuristics: Skeleton-based seeding in Enhanced SIRRT* deterministically initializes tree-based planners along connectivity-representative paths, followed by hybrid smoothing and bidirectional rewiring to propagate structural cost reductions (Ryu, 28 May 2025). Scene-graph-based decomposition, as in S-Path, restricts sampling and divides planning into subproblems aligned to semantic regions, dramatically reducing search complexity (Ejaz et al., 8 Aug 2025).
Graph pattern mining and predictive extension: SPP leverages compression-based subgraph pattern mining and inexact graph matching to infer missing or future semantics, enabling the planner to “look ahead” and adapt plans proactively to the predicted, yet unobserved, structure (Dharmadhikari et al., 6 Jun 2025).
Label-mapping and semantic filtering: Procrustean graphs permit systematic semantic and abstraction-preserving transformations, including union, intersection, and projection (label-maps modeling sensor/actuator degradation), supporting rigorous analysis of plan/filter robustness under representation change (Saberifar et al., 2018).
Semantic aggregation and parallelism: In SPaGe for query-focused summarization, planning over formal step sequences (TaSoF plans) and execution DAGs enables explicit dependency tracking, parallel computation, and reliable translation of queries to executable programs (Zhang et al., 30 Jul 2025).
Retriever-planner decoupling: In symbolic music generation, a structure-aware “style plan” is generated at the section/phrase level by a transformer, then grounded to corpus-pattern retrieval via energy minimization under harmonic, structural, and stylistic constraints (Zang et al., 16 Feb 2026).

4. Practical Applications and Benchmarks

Structure-aware planning frameworks have demonstrated state-of-the-art performance across diverse application domains:

Domain	Key Method(s)	Performance/Impact
Autonomous Driving	SUPER-AD	+9.5 pts EPDMS over no-uncertainty baseline; best DAC
Semantic Inspection	SPP (SPP-AE, SPP-OI)	17–24% faster inspection, 99–99.9% coverage
Human–Robot Assembly	VLM + Symbolic Planning + Minimal Replan	97% state acc., near-zero extra reassignments
Factory Robotics STL	S-MSP (Structured MoE)	+9.5–10.1 pp SR over baselines; improved OOD STL tasks
Manipulator Planning	GAIDE	Lower path cost, higher SR than uniform/neural baselines
UAV Topological Nav.	SphereMap	2–3 orders faster than grid/RRT*; low memory
Table Summarization	SPaGe (Structured Plans)	+3–7 pts BLEU, 98% ESR, 28% cycle reduction
Music Accompaniment	Transformer Style Planner + Retrieval	High diversity, style isolation, realistic performance

Each method’s core advances are traceable to structure encoding: uncertainty maps, symbolic dependency graphs, domain priors, scene graphs, or energy-based retrieval over structural indices (Ryu et al., 28 Nov 2025, Dharmadhikari et al., 6 Jun 2025, Chen et al., 2 Jan 2026, Ye et al., 16 Sep 2025, Soleymanzadeh et al., 3 Mar 2026, Musil et al., 2023, Zhang et al., 30 Jul 2025, Zang et al., 16 Feb 2026).

5. Quantitative Impact and Ablations

Robust empirical evidence supports the structural paradigm:

Safety and robustness: Dense BEV uncertainty and lane regularization in SUPER-AD yield significant gains (+9.5 pts EPDMS, best drivable-area compliance) (Ryu et al., 28 Nov 2025).
Inspection efficiency: In SPP, predictive structure-aware strategies achieve up to 60% faster completion with no loss in coverage, and maintain efficiency under pattern imperfections (Dharmadhikari et al., 6 Jun 2025).
Minimal editing and verification: Structure-aware symbolic plans (SWAP, HPR) maintain plan stability ( $\bar d$ near zero under human intervention) and deliver high-accuracy state synthesis in cluttered conditions (Xiong et al., 2024, Chen et al., 2 Jan 2026).
Compositional task success: MoE specialization in S-MSP yields higher STL satisfaction and generalization compared to single-expert models (Ye et al., 16 Sep 2025).
Sample/compute efficiency: S-Path’s semantic subproblem decomposition produces 5.7x mean planning speedup and up to 52x in replanning cases, while SphereMap achieves real-time multi-goal navigation via topological abstraction and caching (Ejaz et al., 8 Aug 2025, Musil et al., 2023).

Ablation studies consistently show that removing or weakening the structural module (e.g., uncertainty maps, graph structure, plan regularizer) leads to measurable drops in safety, efficiency, and solution quality.

6. Current Limitations and Open Problems

Structure-aware planning faces domain-dependent constraints, including:

Representation complexity: Pattern mining and subgraph-matching (e.g., SPP) have worst-case exponential cost, mitigated in practice by beam search and structure constraints (Dharmadhikari et al., 6 Jun 2025).
Structural misspecification: Overly rigid or inaccurate structural models can lead to efficiency loss or failure in non-conforming scenarios; structure prediction modules (as in SPP or SPaGe) partly mitigate this but hinge on accurate detection or learned priors (Dharmadhikari et al., 6 Jun 2025, Zhang et al., 30 Jul 2025).
Integration with learning: Existing post-hoc safety/verification layers (e.g., TSP repair in S-MSP) could be more tightly integrated with end-to-end differentiable frameworks for improved efficiency and robustness (Ye et al., 16 Sep 2025).
Extension to higher dimensions: Skeleton-based and scene-graph approaches scale poorly to high-DOF or highly unstructured environments; ongoing work includes learning-based abstraction and online reduction techniques (Ryu, 28 May 2025, Soleymanzadeh et al., 3 Mar 2026).
Domain generality: Methods often require domain-specific definitions of structure; cross-domain models for general structural abstraction remain an open challenge (Saberifar et al., 2018).

7. Connections to Theoretical and Symbolic Planning

The language-theoretic foundation for structure-aware planning is formalized in the procrustean graph (p-graph) framework, which generalizes many classical models – finite filters, universal plans, hybrid automata, strategy complexes – into a unified graph abstraction with rigorous semantics in terms of interaction languages (Saberifar et al., 2018). Operations such as union, intersection, label-projection, and normalization enable robust reasoning about plan/filter transformations, abstraction under degradation, and state-equivalence, with provable criteria for behavioral preservation and complexity-theoretic hard limits for certain minimal abstraction tasks. This theoretical framework underpins many practical advances in structure-aware symbolic planning and filtering.

In summary, structure-aware planning constitutes a principled, empirically validated approach for embedding explicit structural information into the planning process, yielding systems that are safer, more efficient, scalable, and better aligned with domain constraints across a wide spectrum of robotics, AI planning, and symbolic reasoning domains (Ryu et al., 28 Nov 2025, Dharmadhikari et al., 6 Jun 2025, Chen et al., 2 Jan 2026, Ye et al., 16 Sep 2025, Soleymanzadeh et al., 3 Mar 2026, Musil et al., 2023, Zhang et al., 30 Jul 2025, Xiong et al., 2024, Zang et al., 16 Feb 2026, Saberifar et al., 2018).