Space Planning Problems (SPP) Overview

• Space Planning Problems (SPP) are a diverse set of optimization and constraint-satisfaction challenges that define feasible spatial, temporal, and resource configurations in domains like architecture, social distancing, and satellite missions.
• SPP formulations integrate explicit constraints such as geometry, adjacency, capacity, and kinematics, employing methods from integer linear programming to reachability analysis and sampling-based algorithms.
• Research in SPP focuses on decomposition strategies and hybrid solution architectures that balance formal guarantees with computational efficiency, enabling practical application in complex, safety-critical environments.

Space Planning Problems (SPP) denote a heterogeneous family of optimization, planning, and constraint-satisfaction problems in which admissible decisions are governed by spatial structure, temporal evolution, resource limits, and safety or feasibility requirements. In the cited literature, the term covers architectural space layout planning, socially distanced seat allocation, general space–time facility activation, safety-critical multi-vehicle path planning, volumetric planning in configuration space, and satellite mission scheduling under physics-based constraints (Medjdoub et al., 2013, Barry et al., 2021, Barbati et al., 2018, Bansal et al., 2016, Shkolnik et al., 2011, Wang et al., 16 Jan 2026). Across these formulations, the common core is the joint selection of entities, locations, geometric relations, and often activation times, subject to exclusion, adjacency, capacity, kinematic, or uncertainty constraints.

1. Terminological scope and problem families

In architectural CAD, Medjdoub & Yannou formulate space layout planning as a constraint-based design problem over rectangular spaces, where the objective is to enumerate consistent conceptual design solutions called topological solutions before introducing exact geometry (Medjdoub et al., 2013). In workplace planning under social distancing, the problem is cast as selecting a maximum number of safe seats under pairwise distance constraints, yielding a binary optimization problem over candidate workspaces (Barry et al., 2021). In operations research, Barbati et al. define a general space–time model whose basic variable $x_{ilt}$ jointly answers what facility is activated, where it is assigned, and when it is started, thereby combining elements of knapsack, location, and scheduling models (Barbati et al., 2018).

A different line of work treats SPP as a dynamic safety problem. In "Safe Sequential Path Planning Under Disturbances and Imperfect Information" (Bansal et al., 2016), the problem is to guarantee collision avoidance and target reachability for multiple vehicles under disturbances and incomplete information. In "Sample-Based Planning with Volumes in Configuration Space" (Shkolnik et al., 2011), planning occurs in $\mathcal C_{\rm free}$, where feasible motion is approximated by convex volumes rather than isolated samples. AstroReason-Bench uses the term in yet another, explicitly space-mission sense: SPP becomes a family of high-stakes planning problems for satellites, ground stations, targets, and inter-satellite links, with heterogeneous objectives and strict physical constraints (Wang et al., 16 Jan 2026).

This distribution of usages suggests that SPP is best understood not as a single canonical problem class, but as a broad modeling idiom centered on spatial feasibility, combinatorial decision structure, and constrained temporal deployment.

2. Core mathematical representations

The architectural formulation begins from a knowledge model in which a Space is an isothetic rectangle with integer-valued, module-based attributes $x_1,y_1,x_2,y_2,L,W,S$ and class-constraints

$x_2=x_1+L,\qquad y_2=y_1+W,\qquad S=L\times W.$

A Room adds orientation $\in\{0^\circ,90^\circ\}$, while a BuildingUnit enforces inclusion in the unit contour, non-overlap among spaces, and contour total-recovery (Medjdoub et al., 2013).

The social-distancing seat-allocation model uses a binary variable

$x_i\in\{0,1\},\quad i\in R,$

with objective

$\max_x \sum_{i\in R} x_i,$

and conflict constraints

$x_i+x_j\le 1,\quad \forall (i,j)\in E_D,$

where $E_D=\{(i,j):i<j,\ d_{ij}<D\}$ and $d_{ij}$ is the Euclidean distance between seat centroids (Barry et al., 2021). This is exactly a maximum independent set problem on the conflict graph $\mathcal C_{\rm free}$0 with $\mathcal C_{\rm free}$1 and $\mathcal C_{\rm free}$2.

The general space–time model takes

$\mathcal C_{\rm free}$3

with $\mathcal C_{\rm free}$4 if facility $\mathcal C_{\rm free}$5 is activated in location $\mathcal C_{\rm free}$6 at the start of period $\mathcal C_{\rm free}$7, and $\mathcal C_{\rm free}$8 otherwise. Its constraints include period budgets, single activation, optional precedence, and optional location/time capacity, while the objective is an aggregated discounted score built from criterion weights, facility-location scores, and discount factors $\mathcal C_{\rm free}$9 (Barbati et al., 2018).

The multi-UAV SPP framework is formulated through Hamilton–Jacobi reachability. For vehicle $x_1,y_1,x_2,y_2,L,W,S$0, the value function satisfies

$x_1,y_1,x_2,y_2,L,W,S$1

for $x_1,y_1,x_2,y_2,L,W,S$2, with terminal condition $x_1,y_1,x_2,y_2,L,W,S$3 (Bansal et al., 2016). The Ball-Tree formulation, by contrast, represents connected subsets of free configuration space with volumes

$x_1,y_1,x_2,y_2,L,W,S$4

and assigns each tree node a center $x_1,y_1,x_2,y_2,L,W,S$5 and radius $x_1,y_1,x_2,y_2,L,W,S$6 (Shkolnik et al., 2011).

AstroReason-Bench defines a space-mission SPP instance by a set of satellites $x_1,y_1,x_2,y_2,L,W,S$7, ground stations $x_1,y_1,x_2,y_2,L,W,S$8, targets $x_1,y_1,x_2,y_2,L,W,S$9, optional inter-satellite links $x_2=x_1+L,\qquad y_2=y_1+W,\qquad S=L\times W.$0, a horizon $x_2=x_1+L,\qquad y_2=y_1+W,\qquad S=L\times W.$1, and decision variables

$x_2=x_1+L,\qquad y_2=y_1+W,\qquad S=L\times W.$2

where each variable indicates scheduling an observation, downlink, slew, or inter-satellite link in a candidate window (Wang et al., 16 Jan 2026).

3. Constraint structures

A defining feature of SPP formulations is the centrality of structured constraints. In architectural layout planning, topological constraints include generalized adjacency$x_2=x_1+L,\qquad y_2=y_1+W,\qquad S=L\times W.$3, on-contour constraints, staircase-entry adjacency, logical disjunctions, and implicit reduction constraints such as non-overlapping$x_2=x_1+L,\qquad y_2=y_1+W,\qquad S=L\times W.$4, incoherent-space-elimination, symmetry elimination among identical spaces, and orientation propagation (Medjdoub et al., 2013). A topological solution is an assignment of all adjacency-direction variables $x_2=x_1+L,\qquad y_2=y_1+W,\qquad S=L\times W.$5, together with orientation choice-points, such that the specification and implicit constraints remain arc-consistent and at least one full geometric instantiation exists.

In the seat-allocation setting, the constraints are pairwise and metric: if two seats are closer than the minimum distance $x_2=x_1+L,\qquad y_2=y_1+W,\qquad S=L\times W.$6, they cannot both be selected. The resulting conflict graph may be decomposed into connected components, and if a component is bipartite, one may choose the larger color class as a feasible allocation for that component (Barry et al., 2021).

The space–time operations-research model enriches spatial placement with budget and temporal structure. Its constraints are

• budget at each period,
• single activation,
• optional precedence,
• optional location/time capacity,
• binary or continuous interpretations of $x_2=x_1+L,\qquad y_2=y_1+W,\qquad S=L\times W.$7, and it also supports multi-stakeholder weights and uncertainty via scenario trees (Barbati et al., 2018).

In safe sequential path planning, the relevant constraints are dynamic and adversarial. Lower-priority vehicles treat higher-priority vehicles as moving obstacles, with three information models: centralized control, least-restrictive control, and robust trajectory tracking. The induced obstacle for a lower-priority vehicle is constructed as an over-approximation of the higher-priority vehicle’s possible positions under the relevant disturbance and information assumptions (Bansal et al., 2016).

AstroReason-Bench organizes constraints into three explicit classes:

• resource constraints, including nonnegative energy buffers and bounded data storage;
• kinematic constraints, requiring enough time between observation windows to accommodate slew and settle times;
• concurrency constraints, limiting simultaneous downlinks and inter-satellite links while allowing communication to overlap observations modulo resource budgets (Wang et al., 16 Jan 2026).

These formulations show that “space” in SPP may refer to Euclidean floorplans, conflict geometry, discrete facility locations, vehicle position tubes, configuration spaces, or orbital opportunity windows, but in each case feasibility is expressed through explicit spatial relations.

4. Decomposition strategies and algorithmic paradigms

A recurrent design principle in SPP research is decomposition. Medjdoub & Yannou explicitly separate topology and geometry: first instantiate directional relations and topological choice-points, then solve the remaining geometric CSP and, if required, optimize it with branch-and-bound (Medjdoub et al., 2013). Their topology enumeration uses a space-selection heuristic of “most constrained first” and a variable-selection heuristic that chooses the adjacency variable with smallest domain for the selected space.

The seat-allocation work similarly exploits problem structure. After building the conflict graph, the graph is decomposed into connected components; bipartite components are handled through two-coloring, while non-bipartite components are modified by deleting nodes participating in odd cycles until bipartiteness is restored (Barry et al., 2021). The same paper compares this graph-partition strategy with a constrained random walk and an ILP formulation.

Sequential path planning achieves a different decomposition: a strict total ordering $x_2=x_1+L,\qquad y_2=y_1+W,\qquad S=L\times W.$8 is imposed on vehicles, so that each vehicle solves a reachability problem only in its own state dimension while treating higher-priority vehicles as time-varying obstacles (Bansal et al., 2016). This converts an otherwise exponentially scaling coupled reachability problem into a linearly scaling sequence of single-vehicle computations.

Shkolnik & Tedrake’s Ball-Tree algorithm decomposes free space into convex volumes rather than points. Expansion uses the nearest-volume rule

$x_2=x_1+L,\qquad y_2=y_1+W,\qquad S=L\times W.$9

and in the inexact variant radii are trimmed whenever collision evidence invalidates part of a ball (Shkolnik et al., 2011). The effect is to reject samples that fall inside already covered volumes and to focus exploration on free-surface regions and narrow passages.

Barbati et al. provide a different kind of decomposition in the objective space. Their model supports weighted-sum optimization, Compromise Programming for balancing location or criterion payoffs, and an interactive methodology based on the Dominance-based Rough Set Approach, in which a decision maker iteratively classifies solutions as “good” or “others” and new inequalities are induced from these judgments (Barbati et al., 2018).

AstroReason-Bench standardizes interaction rather than decomposition of variables. All benchmark families are exposed through a single Model Context Protocol and Python API, with a ReAct loop based on get_state(), stage_action(a), commit(), and evaluate() (Wang et al., 16 Jan 2026). This unification is itself a methodological statement: heterogeneous SPP instances can be given a common action semantics even when their internal objectives differ.

5. Guarantees, complexity, and conservatism

Several SPP formulations emphasize formal guarantees. In architectural planning, the topological enumeration procedure is designed so that no topologies are missed, none are inconsistent, and no duplicates are produced; the worst-case complexity is $\in\{0^\circ,90^\circ\}$0, but in practice the number of topological solutions is reported as less than 100 for $\in\{0^\circ,90^\circ\}$1 (Medjdoub et al., 2013). Once topology is fixed, geometric optimization is solved by depth-first branch-and-bound, with the global constraint Cost < C_1 used to prune inferior solutions.

In safe sequential path planning, safety follows from over-approximating higher-priority vehicles’ reachable positions. Each lower-priority vehicle computes a backward reachable set that excludes

$\in\{0^\circ,90^\circ\}$2

and because each $\in\{0^\circ,90^\circ\}$3 is an over-approximation of all possible higher-priority positions under the adopted assumptions, any real trajectory under the optimal policy avoids collisions in the worst case (Bansal et al., 2016). The complexity drops from exponential in the joint state dimension to $\in\{0^\circ,90^\circ\}$4, where each vehicle solves one BRS PDE and possibly one FRS PDE or one error-bound PDE.

The Ball-Tree method maintains probabilistic completeness in both exact and inexact variants. In the exact case, any uncovered free-surface region has nonzero sampling probability; in the inexact case, repeated probing trims away invalid portions until the same expansion argument applies (Shkolnik et al., 2011). A further $\in\{0^\circ,90^\circ\}$5-Ball-Tree variant guarantees the same completeness reasoning if a feasible path exists whose minimum clearance from $\in\{0^\circ,90^\circ\}$6 exceeds $\in\{0^\circ,90^\circ\}$7.

The seat-allocation literature makes a different trade-off. The constrained random-walk heuristic has no optimality guarantee, whereas the ILP formulation is exact for the binary model and can incorporate additional business rules; the graph-partition approach occupies an intermediate position, exploiting special structure for speed while potentially degrading at larger distancing thresholds (Barry et al., 2021).

These examples illustrate a pervasive tension in SPP research: stronger guarantees typically require stronger modeling assumptions, more conservative over-approximations, or more expensive exact computation.

6. Representative applications and empirical behavior

Empirical studies span design, occupancy planning, autonomous navigation, and satellite operations.

Domain	Representative result	Source
Architectural layout	Two-Floor House: 49 topological solutions in $\in\{0^\circ,90^\circ\}$8 s; geometric optima in $\in\{0^\circ,90^\circ\}$9 s each	(Medjdoub et al., 2013)
Seat allocation	On a 300-seat floorplan at 72 in: Random-walk 149 seats, Graph-partition 162 seats, ILP 162 seats	(Barry et al., 2021)
Multi-UAV path planning	Centralized method: LDT $x_i\in\{0,1\},\quad i\in R,$0; 100% success	(Bansal et al., 2016)
Configuration-space planning	Bug Trap: Ball-Tree mean time 14.6 s vs RRT 75.4 s; mean nodes 246 vs 17 739	(Shkolnik et al., 2011)
Space–time facility planning	Didactic 8-facility / 2-location / 5-year example solved by CPLEX with six activated facilities	(Barbati et al., 2018)
Satellite mission planning	SatNet: $x_i\in\{0,1\},\quad i\in R,$1-MILP achieves $x_i\in\{0,1\},\quad i\in R,$2; best agent listed is Gemini 3 Flash with $x_i\in\{0,1\},\quad i\in R,$3	(Wang et al., 16 Jan 2026)

The architectural case studies include a Maculet House with 72 topologies in 30 min and 72 geometric optima for corridor-area in 33 s, and an Office with Patio with 102 topologies, less than 2 min enumeration, and geometric refinement in less than 2 min (Medjdoub et al., 2013). The social-distancing study reports that the random-walk heuristic is fast but delivers 5–10% fewer seats than the others, while the graph-partition approach is almost as good as the ILP at low–medium distances and degrades by up to 7% at large $x_i\in\{0,1\},\quad i\in R,$4 (Barry et al., 2021).

The UAV study evaluates four vehicles with Dubins-like kinematics

$x_i\in\{0,1\},\quad i\in R,$5

with $x_i\in\{0,1\},\quad i\in R,$6, $x_i\in\{0,1\},\quad i\in R,$7, $x_i\in\{0,1\},\quad i\in R,$8, and $x_i\in\{0,1\},\quad i\in R,$9. Method 2 is the most conservative, with LDT $\max_x \sum_{i\in R} x_i,$0, while Method 3 uses reduced turn-rate, fixed speed, and error bound $\max_x \sum_{i\in R} x_i,$1, yielding LDT $\max_x \sum_{i\in R} x_i,$2 (Bansal et al., 2016).

AstroReason-Bench broadens empirical SPP evaluation to agentic systems. Across five benchmark families, current agents substantially underperform specialized solvers on some tasks while showing nontrivial competence on others. For Revisit Optimization, SA achieves $\max_x \sum_{i\in R} x_i,$3 h and Claude Sonnet 4.5 achieves $\max_x \sum_{i\in R} x_i,$4 h with $\max_x \sum_{i\in R} x_i,$5; for Regional Coverage, Gemini 3 Flash attains $\max_x \sum_{i\in R} x_i,$6; for Stereo Imaging, Qwen3 Coder reaches $\max_x \sum_{i\in R} x_i,$7; and for Latency Optimization, Kat Coder Pro is the only listed agent with nonzero availability, $\max_x \sum_{i\in R} x_i,$8 and $\max_x \sum_{i\in R} x_i,$9 ms (Wang et al., 16 Jan 2026).

7. Limitations, misconceptions, and current research directions

A common misconception is that SPP denotes one stable mathematical object. The surveyed literature shows instead that the label spans several problem classes whose unifying feature is constrained spatial decision-making rather than a single shared formalism. This suggests caution when transferring algorithms or benchmark claims across subfields.

The limitations are domain-specific. The architectural framework is restricted to orthogonal, isothetic rectangles; corridors are limited to at most two rectangles; disjunctions of topological specifications are handled only in restricted form; and optimization criteria are currently limited to corridor area and wall length (Medjdoub et al., 2013). In multi-UAV planning, the least-restrictive control model is deliberately conservative because lower-priority vehicles only know the intersection of a higher-priority vehicle’s BRS and FRS, not its realized policy (Bansal et al., 2016). In Ball-Tree planning, the exact version requires clearance queries that may be expensive in high-DOF or complex geometry, and in obstacle-dense regions the method can collapse toward standard RRT behavior because radii shrink toward zero (Shkolnik et al., 2011). In seat allocation, heuristic methods trade optimality for speed, and larger social-distance thresholds can erode the quality of graph-partition solutions (Barry et al., 2021).

The current frontier in space-mission SPP emphasizes benchmarking and hybridization. AstroReason-Bench identifies resource mismanagement, geometric ignorance, insufficient environment exploration, and failures of compound constraint reasoning as recurring agent failure modes, while also proposing directions such as hybrid LLM+MILP systems, explicit search/planning submodules, analysis-versus-action phases, multi-agent training, deep-space trajectory design, satellite constellation design, and architectural co-optimization (Wang et al., 16 Jan 2026). Barbati et al. point toward uncertain and multi-stakeholder planning through scenario trees, Compromise Programming, and DRSA, while Medjdoub & Yannou propose polygonal or curved building contours, richer functional specifications, uncertain or soft constraints with weighted penalties, multi-floor stacked units, and industrial plant layout extensions (Barbati et al., 2018, Medjdoub et al., 2013).

Taken together, these directions indicate that SPP research is moving toward richer uncertainty models, stronger interaction between geometry and combinatorics, and hybrid solver architectures that combine exact optimization, reachability, sampling, and agentic planning under a common spatial decision framework.