Prior-Guided Path Planning

• Prior-guided path planning is an architectural approach that uses diverse prior information—such as natural-language instructions, waypoint sequences, and learned feasibility maps—to inform and restrict path search.
• It integrates various guidance signals to reduce the search space and computational complexity while ensuring routes are safe, feasible, and aligned with task-specific constraints.
• Despite its benefits, challenges remain in managing conflicting priors, computational overhead, and ensuring adaptability across different dynamic and high-dimensional scenarios.

Prior-guided path planning denotes a class of planning methods in which the planner is not driven solely by free-space geometry and obstacle avoidance, but is instead conditioned by additional information available before or during search. In recent work, that guiding information has taken the form of natural-language instructions, precomputed global paths and waypoint sequences, curvature-feasible motion primitives, learned soft path fields, previously reconstructed models, stored databases of prior paths, latent priors for diffusion-based trajectory generation, topological classes, and explicit task-order constraints such as key-door precedence (Bao et al., 23 Feb 2025, Guvenkaya et al., 11 Nov 2025, Schichler et al., 2024, Kim et al., 8 Jun 2026, Attali et al., 7 Apr 2025, 2505.10881, Kim et al., 2021, You et al., 24 Oct 2025). Across these formulations, the prior does not replace planning; it biases cost construction, search expansion, sampling, graph augmentation, or trajectory generation so that planning is directed toward routes that are feasible, safe, task-aligned, or computationally easier to extract.

1. Operational scope of the “prior”

In this literature, the word prior is used in several technically distinct ways. Some papers use it as a conditioning signal over route preference, as in instruction-conditioned cost prediction for roadmaps (Bao et al., 23 Feb 2025). Others use it as precomputed structural guidance, such as a global path reduced to waypoints for local obstacle avoidance (Guvenkaya et al., 11 Nov 2025), or a library of admissible motion primitives that already encode curvature feasibility (Schichler et al., 2024). A different line treats the prior as a learned spatial template over feasible paths, which is then adapted to new objectives rather than relearned from scratch (Kim et al., 8 Jun 2026). Experience-based planners interpret the prior as a database of solved paths that can be queried online as a heuristic memory (Attali et al., 7 Apr 2025). Temporal reconstruction work uses the previous cycle’s plant model as a plant-specific temporal prior for current-cycle view planning (Pan et al., 8 Oct 2025). Task-logic planners encode prior knowledge about precedence directly into an augmented graph structure (You et al., 24 Oct 2025).

Prior source	Planning role	Representative papers
Natural-language instruction	Conditioned cost-map prediction and roadmap bias	(Bao et al., 23 Feb 2025)
Global path and waypoints	Constrain local planning between consecutive waypoints	(Guvenkaya et al., 11 Nov 2025)
Motion primitives	Restrict graph expansion to curvature-feasible cell sequences	(Schichler et al., 2024)
Learned soft path field	Reusable feasibility prior for objective-level adaptation	(Kim et al., 8 Jun 2026)
Path database	Heuristic cost-to-go with online path-suffix updates	(Attali et al., 7 Apr 2025)
Previous reconstruction	Temporal prior for minimal view-set planning	(Pan et al., 8 Oct 2025)
Precedence specification	Layered graph augmentation by accumulated key sets	(You et al., 24 Oct 2025)

This diversity suggests that prior-guided planning is best understood as an architectural pattern rather than a single algorithmic family. The shared feature is that the prior alters the geometry of planning before final path extraction. In some cases the prior changes the cost landscape; in others it changes the graph, the admissible successor set, the sampling distribution, or the state representation itself.

A recurring misconception is that prior-guided planning is necessarily synonymous with learned planning. The surveyed papers do not support that equivalence. Topological priors derived from a generalized Voronoi diagram, global waypoint priors, motion-primitive priors, and key-door precedence priors are all non-learned, yet they still function as strong planning priors (Kim et al., 2021, Guvenkaya et al., 11 Nov 2025, Schichler et al., 2024, You et al., 24 Oct 2025). Conversely, learned priors often remain tightly coupled to classical search rather than replacing it outright, as in roadmap search, Focal Search, Multi-Heuristic A*, or shortest-path solvers (Bao et al., 23 Feb 2025, Kim et al., 8 Jun 2026).

2. Search-space restriction, graph bias, and task structure

A central role of priors is to reduce the effective search space before expensive optimization begins. In curvature-aware autonomous-vehicle planning, a modified A* searches on a hexagonal grid but only expands cell transitions that match a finite set of admissible motion primitives of length $n=5$. The total search cost is written as

$C = C_c + C_g,$

with

$C_c = W_n \cdot n^* + W_{CK} C_K.$

Here the prior is the primitive library itself: it encodes local path formations compatible with the vehicle’s minimum turning radius, and the curvature term $C_K$ is derived from the curvature of a smooth path embedded within the primitive rather than from raw grid turning angle (Schichler et al., 2024). The result is that A* no longer explores arbitrary neighboring cells; it explores only locally feasible templates.

Topology-guided visual navigation for MAVs applies the same logic at a different structural level. A 2D generalized Voronoi diagram is extracted from the ESDF, converted into a sparse topological graph, and used to enumerate initial paths in different homology classes. Each candidate path is recursively segmented according to the co-visibility ratio

$$CR(\mathbf{p}, \mathbf{q}) = \frac{|CV(\mathbf{p}, \mathbf{q})|}{|V(\mathbf{p}) \cup V(\mathbf{q})|},$$

and within each class a layered 4-DoF pose graph is optimized under

$$\mathrm{P}^* = \arg\min_{k_1,\ldots,k_N} \lambda_d c_d(\mathrm{P}) - \lambda_p c_p(\mathrm{P}).$$

The topological prior is therefore not a final path but a finite family of obstacle-relative classes within which perception quality is evaluated (Kim et al., 2021).

Task-order priors can be even more explicit. In motion planning with key-door precedence, the ordering constraint is written as a fragment of signal temporal logic, for example

$$\varphi = (\neg \mathcal{D}_1\,\mathcal{U}\,\mathcal{K}_1)\wedge \cdots \wedge (\neg \mathcal{D}_n\,\mathcal{U}\,\mathcal{K}_n)\wedge \mathcal{F}(\mathcal{T}).$$

The method constructs an exact convex partition of free space from a hyperplane arrangement, labels convex cells as free space, key sets, or door sets, and then builds an augmented graph of convex sets whose layers correspond to reachable subsets of collected keys. Doors are inaccessible in the base layer and are re-enabled only in layers where the corresponding keys have been acquired (You et al., 24 Oct 2025). The prior is therefore embedded directly into graph topology.

Subgoal-based planning for autonomous vehicles provides a related but control-theoretic interpretation. Candidate subgoals are defined on obstacle boundaries through constrained optimal control conditions, with admissible candidates

$$\mathcal{G}(\mathbf{x}_g, \pi, O_i) := S_B(\mathbf{x}_g, \pi, O_i) \cap \partial O_i.$$

Planning then becomes a graph search over meaningful transition states rather than a search over arbitrary samples. This reduces the continuous problem to a subgoal sequence that links reusable motion-guidance elements into a constrained trajectory (Feit et al., 2020).

Local planning with dynamic obstacles also relies on structural prior. The robot assumes that a global path has already been computed with any well-known algorithm, such as A* or RRT*, and that extracted waypoints are provided to the robot. The local planner then operates only between consecutive waypoints, classifies the current situation into six cases, and combines tangent-based bypassing with extrapolation-based collision-time reasoning (Guvenkaya et al., 11 Nov 2025). In this setting, the prior is the precomputed global route, which constrains local decisions to a short, directed segment.

3. Learned semantic, language-conditioned, and foundation-model priors

A major contemporary direction is to learn a reusable prior from data and then use classical planning on top of it. Instruction-Guided Probabilistic Roadmaps exemplifies the direct use of language as a route-preference prior. The continuous planning problem is posed over $X=[0,1]\subseteq \mathbb{R}^d$, with feasible paths $\sigma:[0,1]\to X$, but the cost map $C = C_c + C_g,$0 is no longer fixed a priori. Instead, a natural-language instruction is embedded by an off-the-shelf OpenAI model such as text-embedding-ada-002 or text-embedding-3-small, producing a 1536-dimensional vector that is reduced to 128 dimensions using Gaussian random projection. The reduced embedding is concatenated with the occupancy map and fed to a U-Net with a VGG-16 backbone to predict an instruction-guided cost map $C = C_c + C_g,$1. The roadmap then samples nodes with probability inversely proportional to the predicted cost, computes edge costs along line segments through the predicted map, and runs Dijkstra search (Bao et al., 23 Feb 2025). The prior acts twice: in cost prediction and in roadmap construction.

FlexPath separates feasibility from preference more explicitly. Stage 1 learns a task-independent spatial prior from rasterized semantic grids $C = C_c + C_g,$2, producing a dense path field $C = C_c + C_g,$3 and a soft mask

$C = C_c + C_g,$4

This prior is trained by imitation learning with a recall-biased weighted binary cross-entropy that emphasizes false negatives and yields a “thicker” connectivity-preserving path prior. Stage 2 does not relearn path structure; it adapts $C = C_c + C_g,$5 using differentiable Path Shape Objectives for shortest path, obstacle clearance, semantic avoidance, or waypoint following, after which classical planners still extract the discrete trajectory (Kim et al., 8 Jun 2026). This design formalizes the idea that prior-guided planning can reuse a single learned feasibility prior across multiple downstream objectives.

Foundation-model guidance pushes the semantic prior to a higher abstraction level. FM-Planner supplies a tokenized prompt

$C = C_c + C_g,$6

to an LLM or VLM and asks for a sparse global waypoint sequence. Its practical deployment variant uses YOLOv8 to detect obstacles and estimate their positions, serializes the detected objects and scene labels into text, and feeds that representation to a fine-tuned LLM. The generated waypoint list is then interpolated, with a stated safety margin of $C = C_c + C_g,$7 from obstacles in the online stage (Xiao et al., 27 May 2025). The prior here is high-level semantic reasoning over structured scene descriptions rather than a conventional geometric heuristic.

A reinforcement-learning formulation for UAV weed search uses a different semantic prior: a global map encoding uncertain prior knowledge about object locations, combined with a local map of in-flight detections and a movement budget scalar. The policy is learned with Deep Q-Learning, using two parallel convolutional branches for the global prior map and local detection map. This prior is not assumed to be highly accurate; the experiments explicitly corrupt it with false positives, false negatives, positional noise, and downsampling (Essen et al., 2024). The planner thus combines weak long-range guidance with stronger local evidence.

These learned methods collectively show that priors can encode preferences that are not reducible to shortest-path geometry: wider passages, semantic avoidance, waypoint adherence, or clustered target search. They also show that the prior often remains a soft field or coarse waypoint structure rather than a complete solution.

4. Experience, temporal carry-over, and adaptive priors

Experience-based and temporally propagated priors occupy a distinct part of the design space. Path Database Guidance stores solutions to previously seen motion-planning problems in a database $C = C_c + C_g,$8, but does not simply retrieve and paste a path. Instead, it computes a path-closest-value

$C = C_c + C_g,$9

and then a database value

$C_c = W_n \cdot n^* + W_{CK} C_K.$0

with the practical local version

$C_c = W_n \cdot n^* + W_{CK} C_K.$1

Whenever an attachment edge is invalidated by collision checking, the method deletes the invalidated prefix of the stored path and retains only the suffix for future guidance (Attali et al., 7 Apr 2025). The database is therefore not treated as a static prior; it is rewritten during search.

Temporal-prior-guided view planning for periodic 3D plant reconstruction uses a previously reconstructed model of the same plant as a plant-specific temporal prior. The previous-cycle complete point cloud $C_c = W_n \cdot n^* + W_{CK} C_K.$2 is non-rigidly aligned to the current partial observation $C_c = W_n \cdot n^* + W_{CK} C_K.$3 by embedded deformation,

$C_c = W_n \cdot n^* + W_{CK} C_K.$4

optimized with

$C_c = W_n \cdot n^* + W_{CK} C_K.$5

using $C_c = W_n \cdot n^* + W_{CK} C_K.$6. The aligned model is inflated to account for growth and uncertainty, visible surface points are extracted with OctoMap, and an integer linear program selects a minimal set of additional views: $C_c = W_n \cdot n^* + W_{CK} C_K.$7 subject to coverage constraints over all surface points (Pan et al., 8 Oct 2025). The final viewing order is then computed as a globally shortest path over the selected views.

Hybrid planners also use prior paths as initialization for local geometric optimization. In smooth path planning with a Gaussian process regression map, weighted A* or T-RRT first produces a coarse path on a continuous differentiable map of traversability, obstacle distance, and posterior variance. That path initializes the control points of a Bézier curve, which is then optimized under

$C_c = W_n \cdot n^* + W_{CK} C_K.$8

BCO-$C_c = W_n \cdot n^* + W_{CK} C_K.$9 initializes control points on the straight line from start to goal, whereas BCO-A* and BCO-RRT place them along the prior path, which gives the optimizer a better starting point and helps preserve a valid route class (Serdel et al., 2024).

Adaptive motion planning with artificial potential fields uses a prior path in a similarly explicit way. A feasible coarse path is obtained by RRT or another planner, then turned into a directive force

$C_K$0

where the active segment is chosen by the distance criterion

$C_K$1

The sampling space for generating later prior paths is also reduced by

$C_K$2

The prior path breaks the journey into intermediate sub-goals and is then combined with APF reactivity and genetic-algorithm tuning of force weights (Amiryan et al., 2020).

In unknown environments, even “currently available information” can function as a prior. The optimal-transport-inspired graph planner begins from the initial known obstacle set, the target-centered potential $C_K$3, and previously detected obstacles. Its graph is rebuilt as newly sensed obstacles are added to the known environment, and the explored sub-region is characterized through a Fokker–Planck-inspired construction (Zhai et al., 2019). This suggests that prior-guided planning need not presuppose a fully known environment; it can also mean planning from a progressively updated partial model.

5. Dynamics, safety, and embodiment-aware priors

When planning moves from geometric routing to full trajectories, the prior often acquires dynamical or safety content. In offline reinforcement learning, Prior Guidance replaces the standard Gaussian diffusion prior

$C_K$4

with a learnable conditional prior

$C_K$5

while keeping the denoiser $C_K$6 fixed. The behavior-regularized latent objective is

$C_K$7

The prior is therefore a distribution over latent noise that decodes into higher-value trajectories, rather than an explicit geometric map or graph bias (2505.10881).

GADGET implements a closely related idea for manipulator planning but combines several priors simultaneously. A conditional diffusion model generates joint-space trajectories conditioned on start configuration, goal configuration, and a learned voxelized scene embedding $C_K$8. Classifier-free guidance strengthens scene adherence, while classifier-guided Control Barrier Functions add a safety prior during denoising. For link $C_K$9 and obstacle $$CR(\mathbf{p}, \mathbf{q}) = \frac{|CV(\mathbf{p}, \mathbf{q})|}{|V(\mathbf{p}) \cup V(\mathbf{q})|},$$0, the barrier is

$$CR(\mathbf{p}, \mathbf{q}) = \frac{|CV(\mathbf{p}, \mathbf{q})|}{|V(\mathbf{p}) \cup V(\mathbf{q})|},$$1

and the soft penalty is

$$CR(\mathbf{p}, \mathbf{q}) = \frac{|CV(\mathbf{p}, \mathbf{q})|}{|V(\mathbf{p}) \cup V(\mathbf{q})|},$$2

The denoising step is then corrected by $$CR(\mathbf{p}, \mathbf{q}) = \frac{|CV(\mathbf{p}, \mathbf{q})|}{|V(\mathbf{p}) \cup V(\mathbf{q})|},$$3 (Tamizi et al., 21 Oct 2025). Here the prior is simultaneously learned, scene-aware, and safety-shaped.

For high-dimensional walking robots, prior guidance becomes a practical necessity because feasibility checks are expensive. Informed Guided RRT*-Connect first computes a 2D A* guide path on a simplified map, uses that guide to place temporary goals for RRT-Connect in the full robot state space, and then injects the resulting feasible path into an informed RRT*-Connect optimization stage that samples inside the ellipsoid defined by start, goal, and current path cost (Belter, 2023). The planner also uses a posture optimization objective

$$CR(\mathbf{p}, \mathbf{q}) = \frac{|CV(\mathbf{p}, \mathbf{q})|}{|V(\mathbf{p}) \cup V(\mathbf{q})|},$$4

The guiding prior is thus a coarse global route that makes expensive full-body sampling tractable.

These trajectory-level examples show that prior-guided planning is not confined to 2D or grid-based settings. The prior may instead encode desirable latent regions, embodiment-aware safety structure, or coarse global guidance for high-dimensional constrained motion.

6. Empirical tendencies, limitations, and open questions

The reported empirical pattern across these papers is that priors improve either efficiency, task alignment, safety, or all three, but usually not without qualification. On synthetic maps and IN2D indoor maps, IG-PRM achieves higher SPL and lower DTW than standard PRM for node budgets of 50, 150, and 300, and generalizes reasonably well to unseen instructions, although performance is better on known instructions (Bao et al., 23 Feb 2025). On TMP 640k, FlexPath reduces search effort on shortest-path planning by 14.3% compared to TransPath, and for obstacle clearance with minimum clearance distance 2 it achieves 96.8% full obstacle avoidance while maintaining low search cost; for waypoint following it succeeds in 98.3% of test cases (Kim et al., 8 Jun 2026). In periodic plant reconstruction, the temporal-prior-guided pipeline maintains or improves surface coverage while requiring fewer views and comparable movement cost compared to the state-of-the-art baselines, and adding one NBV before registration reduces registration Chamfer distance from $$CR(\mathbf{p}, \mathbf{q}) = \frac{|CV(\mathbf{p}, \mathbf{q})|}{|V(\mathbf{p}) \cup V(\mathbf{q})|},$$5 mm to $$CR(\mathbf{p}, \mathbf{q}) = \frac{|CV(\mathbf{p}, \mathbf{q})|}{|V(\mathbf{p}) \cup V(\mathbf{q})|},$$6 mm (Pan et al., 8 Oct 2025).

Real-time and fielded evaluations exhibit the same pattern of benefit with caveats. In local path planning for UGVs with dynamic obstacles, success ratios are consistently above 85% at low density, fall to about 70%–90% at medium density, and drop substantially at high density, with a 3 m critical area radius giving the best balanced performance overall (Guvenkaya et al., 11 Nov 2025). In UAV weed localization, the learned policy is especially effective when weeds are non-uniformly distributed, and in real-world transfer it produced a path about 66% shorter than a row-by-row baseline at the cost of a 10% lower percentage of found weeds (Essen et al., 2024). FM-Planner reports that Llama-3.1-8B-Instruct achieved 100% SR in the 2-obstacle and 3-obstacle cases, while the integrated LLM-Vision system succeeded in all three reported real-world drone tasks, with reasoning time about 9–10 s per query (Xiao et al., 27 May 2025). GADGET with CBF guidance improves over the version without CBF in spherical-obstacle, bin-picking, and shelf environments, and physical execution on a Kinova Gen3 produced 38 successful trajectories out of 50 sampled, 30 of which were collision-free in real execution (Tamizi et al., 21 Oct 2025).

The limitations are equally consistent. Instruction guidance can hallucinate: IG-PRM cannot determine whether an instruction is actually relevant to the task, will still produce some cost map for unrelated text, and cannot properly handle contradictory multiple instructions (Bao et al., 23 Feb 2025). Learned priors alone may not encode the downstream preference of interest: FlexPath’s PT prior alone has zero full avoidance for obstacle-clearance planning at $$CR(\mathbf{p}, \mathbf{q}) = \frac{|CV(\mathbf{p}, \mathbf{q})|}{|V(\mathbf{p}) \cup V(\mathbf{q})|},$$7, which is why Stage 2 objective adaptation is necessary (Kim et al., 8 Jun 2026). Some priors are computationally favorable only within a structured regime. The augmented graph of convex sets for key-door precedence has a worst-case upper bound

$$CR(\mathbf{p}, \mathbf{q}) = \frac{|CV(\mathbf{p}, \mathbf{q})|}{|V(\mathbf{p}) \cup V(\mathbf{q})|},$$8

so explicit task-logic augmentation remains exponential in the number of keys in the worst case (You et al., 24 Oct 2025). Cross-embodiment diffusion transfer works best for manipulators with similar serial-chain kinematics, and real-world performance can degrade because of imperfect reconstruction and digital-twin mismatch (Tamizi et al., 21 Oct 2025).

Three broader conclusions follow from the surveyed work. First, prior-guided planning does not eliminate classical planning; it typically reshapes it. Dijkstra search, Focal Search, Multi-Heuristic A*, RRT variants, Hamiltonian-path sequencing, and graph backtracking remain central components (Bao et al., 23 Feb 2025, Kim et al., 8 Jun 2026, Attali et al., 7 Apr 2025, Pan et al., 8 Oct 2025). Second, the strongest gains often arise when feasibility and preference are separated, so that a reusable prior is learned or constructed once and adapted later (Kim et al., 8 Jun 2026, 2505.10881). Third, a plausible implication is that future work will increasingly focus on priors that are simultaneously transferable, safety-aware, and updateable online, because the current literature repeatedly identifies hallucination, objective mismatch, domain shift, and static-prior assumptions as the main bottlenecks.