Visibility Graph Inspired Path Planning (VGIPP)

• Visibility Graph Inspired Path Planning (VGIPP) is a family of algorithms that use line-of-sight relationships among obstacles to compute globally optimal, collision-free paths in both 2D and 3D spaces.
• Dynamic and layered graph structures integrate real-time sensor data and environmental changes, enabling rapid replanning and efficient trajectory optimization.
• Extensions to grid-based and high-dimensional cases, including rotation-stacked graphs, provide scalable solutions with strong empirical optimality and practical handling of nonholonomic constraints.

Visibility Graph Inspired Path Planning (VGIPP) is a family of geometric and graph-theoretic algorithms that leverage the explicit or implicit computation of line-of-sight (visibility) relationships among free-space-critical points (typically obstacle polygon vertices and query poses) to enable rapid, global, and often optimal path planning in 2D and 3D environments. Classical visibility graphs have been extended and adapted to dynamic, partially known, high-dimensional, and grid-discretized domains in both research and applications. VGIPP methods, by explicitly modeling sensory constraints, environmental geometry, and real-time updates, remain foundational for robotic navigation, multi-query motion planning, as well as grid-based search and post-processing routines.

1. Mathematical Foundations and Classical Formulation

For a robot navigating a 2D workspace $\mathcal{Q}\subset\mathbb{R}^2$ with non-overlapping, closed, polygonal obstacles $\{P_i\}_{i=1}^m$, the free space is defined as $$\mathcal{F} = \mathcal{Q} \setminus \bigcup_{i=1}^m \mathrm{Int}(P_i)$$. The canonical visibility graph $G=(V,E)$ is constructed by letting $V$ comprise all obstacle vertices as well as the start and goal points. An undirected edge $(u,v)$ exists if the open line segment $\overline{uv}\subset\mathcal{F}$; the edge weight is the Euclidean length $c(e_{uv}) = \|u-v\|_2$. Computing the shortest path then reduces to a graph search over $G$ under cost $c$, guaranteeing the globally minimum-length, piecewise-linear, collision-free path in polygonal free space (Yang et al., 2021).

The complexity of naive edge tests is $O(n^2 m)$ (with $n$ total vertices), though plane-sweep and output-sensitive algorithms can reduce this to $O(n^2\log n)$ (Yang et al., 2021).

2. Dynamic and Layered Visibility Graphs in Partial and Unknown Environments

Modern VGIPP approaches extend this framework via dynamic, sensor-driven updates and hierarchical data structures. The FAR Planner, for example, utilizes a two-layer architecture:

• Local Layer: Built anew at each time step from current sensor data, capturing recent point clouds and extracting obstacle polygons within a fixed robot-centered window.
• Global Layer: Accumulates and maintains a global map of obstacle polygons and visibility edges across all encountered frames.

A dynamic update algorithm merges the local graph into the global layer by associating new and existing vertices (using geometric proximity and robust fitting, e.g., RANSAC) and pruning occluded or spurious edges. When environmental changes, such as new obstacle detection or path blockage, are observed, rapid replanning is triggered using Dijkstra/A* on the sparsified visibility graph. Planning latency is typically below $10$ ms per replan on commodity CPUs for local vertex sizes $n_\ell \lesssim 100$ and global graphs $|V_g|\approx 200$ (Yang et al., 2021).

Attemptable planning refers to interleaving exploration of both observed free space and as-yet-unknown regions. If no collision-free path is available in the currently mapped free space, the planner is permitted to tentatively plan into unknown space—executing up to the farthest-known-unknown transition, observing further, and replanning as new information arrives.

3. High-Dimensional and Holonomic Extensions: Rotation-Stacked Visibility Graphs

VGIPP principles extend naturally to higher-dimensional and more complex robots. In "Asymptotically-Optimal Multi-Query Path Planning for a Polygonal Robot" (Zhang et al., 2024), the configuration space $\mathcal{C}_{\mathrm{free}}\subset SE(2)$ for a rigid, translating-and-rotating polygonal robot is constructed by Minkowski-growing obstacles with the reflected robot footprint. Global optimality can be retained by discretizing orientation into $m$ intervals, building a reduced visibility graph $\mathcal{VG}_i$ for each slice, and "stacking" them into a rotation-stacked graph (RVG).

Adjacent slices are connected via inter-layer edges between coincident, mutually-visible reflex vertices, permitting transitions in orientation. Query configurations (start and goal) are inserted into all intersecting orientation layers, then connected to visible graph vertices. Running A* (with appropriate translation and rotation-weighted metrics) on the RVG yields a path that is both resolution complete and asymptotically optimal as $m\to\infty$.

Preprocessing (Minkowski sums, graph construction across $m$ slices) is $O(m n^2 k)$ for $n$ total vertices and $k$ obstacles. Once constructed, queries are answered in $<0.05$ s, with path costs within $2$–$3\%$ of the continuous optimum. Path quality increases monotonically with finer orientation discretization, with diminishing returns beyond $m \approx 90$ (Zhang et al., 2024).

4. VGIPP on Grids: Scalar-Field and Implicit Visibility Methods

Extensions of VGIPP to grid-based search exploit fast visibility field computation for both explicit and implicit visibility graphs:

• Hyperbolic PDE Visibility Field: The method in (Ibrahim et al., 2024) propagates a "visibility quantity" $u(x,y)$ over the entire grid via a linear first-order hyperbolic PDE, with an entropy-satisfying upwind scheme in $O(n)$ time. The resulting field allows global visibility from a source to all cells, enabling both heuristic evaluation and deterministic, local-minima-free search.
• Exact Wavefront Propagation (VBM): The algorithm in (Ibrahim et al., 2024) conducts an implicit wavefront using analytic distance pivoting—distance from a pivot propagates to all visible cells, with each cell's anchor updated as needed. Only neighbor lines-of-sight are checked using a dynamic-programming visibility DP. Complexity for wavefront propagation is $O(n)$ (each pivot's visibility field built in a single raster pass and stored as a bitset hash), delivering globally optimal any-angle paths.

Both approaches outperform classical ray-casting and standard marching (e.g., multi-stencil fast marching), especially in cluttered environments, and can act as components in hybrid grid-VGIPP pipelines.

5. VGIPP for Post-Processing and Hybrid Approaches

Homotopy-constrained post-processing leverages visibility graphs to optimally "shortcut" grid paths. The Homotopic Visibility Graph (HVG) method in (Kumar et al., 2021) constructs local subgraphs of visible convex obstacle corners around the initial A*-planned grid path, using orthogonal ray-casting. By searching this pruned graph (with edges given by unobstructed line segments), HVG produces a path that is guaranteed at least as short as any feasible path in the same topological class. It scales efficiently for large maps, with parallelizable ray-casting and $O(k^2 L)$ complexity for $k$ candidate corners and $L$ the cost of a LOS test.

Pruned VGIPP approaches such as FA-A* (Cao et al., 2017) focus only on "candidate" obstacle vertices essential for covering lines of sight between the current node and the goal, thereby maintaining both optimality and computational efficiency, especially for sparse or clustered environments.

6. Augmented Visibility Graphs with Kinematic and Dynamic Constraints

Fielded robotics applications frequently revisit the VGIPP paradigm but modify cost structures to better accommodate system dynamics:

• Dynamic Augmented Visibility Graph (DAVG) (Hou et al., 3 Apr 2025) activates only a minimal bounding region of obstacles for a given planning query. States are augmented with turning angles, enforcing an edge-dependent cost composed of Euclidean translation and penalized absolute turning angles: $w = d_{\mathrm{segment}} + \lambda |\Delta \theta|$. A variant of Dijkstra/A* search on this lifted graph identifies time-efficient, smooth, turn-aware trajectories, which are subsequently interpolated for a continuous reference path.
• In closed-loop frameworks, tracking is ensured with model predictive control (cf-MPC), enforcing collision constraints and robot dynamics, yielding seamless switching between global re-routing (on DAVG) and reactive local control.

This layered architecture supports real-time performance in dynamic, adversarial, and high-noise settings, with reported control loop rates of 120–400 Hz on standard multicore CPUs (Hou et al., 3 Apr 2025).

7. Computational Performance, Scalability, and Theoretical Guarantees

VGIPP methods deliver strong theoretical and empirical guarantees:

• Optimality: Classical VG, RVG, HVG, and related approaches yield globally shortest or homotopically optimal paths under piecewise-linear and Euclidean metrics. Resolution completeness and asymptotic optimality have been formally established for rotation-stacked visibility graphs as discretization is refined (Zhang et al., 2024).
• Computational Complexity: With appropriate pruning, layering, and deferred visibility computation, VGIPP runtimes can match or surpass sampling-based and dense-grid planners in both preprocessing and per-query time, maintaining $O(n^2 \log n)$ to $O(n)$ asymptotics depending on representation and update regime.
• Empirical Results: FAR Planner demonstrates $12$–$47\%$ shorter paths than A*, D* Lite, and $24$–$35\%$ over RRT*, BIT*, SPARS in representative environments; planning latency is consistently below $10$ ms (Yang et al., 2021). Grid-embedded VGIPP planners either match or improve upon any-angle path cost and runtime baselines such as Anya, Theta*, or MSFM (Ibrahim et al., 2024, Ibrahim et al., 2024).

Limitations include scalability challenges for high numbers of vertices in dense and high-dimensional environments, increased memory with orientation slicing, and the requirement for specialized handling for nonholonomic or kinodynamically-constrained systems.

8. Extensions and Ongoing Directions

VGIPP continues to serve as a touchstone for further innovations:

• 3D and Hybrid Representations: By extending slices to multiple spatial layers and by interconnecting visibility across these strata, VGIPP approaches can generalize to 3D (e.g., for aerial or underwater navigation) (Yang et al., 2021).
• Learning and Adaptivity: Incorporating learned or data-driven heuristics into VGIPP edge or node costings can bias search efficiently in unknown regions.
• Parallelization and Hardware Acceleration: Many components of VGIPP pipelines (visibility computation, ray-casting, convex hull pruning) are highly parallelizable and amenable to SIMD/GPU acceleration.
• Adaptive Discretization: Concentrating orientation or spatial slices in environments with narrow passages or near critical features can reduce overall complexity while maintaining path quality (Zhang et al., 2024).
• Kinodynamic and Nonholonomic Integration: Edges can be augmented or filtered using reachability/constrained feasibility tests (e.g., Reeds-Shepp for car-like vehicles), or obstacle inflation to represent minimum-turn-radius feasibility (Yang et al., 2021).

VGIPP retains favored status for tasks where global path efficiency, rapid replanning, and geometric exactness are paramount, and continues to inspire hybrid systems unifying geometric, grid, and learning-based motion planning.