GEGRH: Multi-Hypothesis Navigation Planning

• GEGRH is a kinodynamic multi-hypothesis planning strategy that defers expensive cross-hypothesis rerouting until goal-edge expansion.
• The method incorporates graph revision to redirect focus from goal-centric expansions to divergence points, improving search efficiency.
• GEGRH balances speed and safety by averaging costs across conflicting world models to yield robust mobile robot navigation.

Searching arXiv for the specified paper and closely related motion-planning context. GEGRH, short for Goal-Edge Graph Revision Hypothesis, is a kinodynamic graph-search method for mobile robot navigation under inconsistent world models. It addresses the case in which successive cost maps produced from onboard perception can flip regions between obstacle and free space, causing deterministic planners to alternate between topologically distinct trajectories across planning cycles. In the formulation reported in "Kinodynamic Motion Planning for Mobile Robot Navigation across Inconsistent World Models" (Damm et al., 30 Sep 2025), GEGRH reasons over a short history of hypotheses rather than relying only on the latest map, but it does not require every candidate edge to be valid in all hypotheses. Instead, it defers hypothesis-specific rerouting until a goal-edge expansion occurs, then performs a graph revision step that shifts search effort toward the earliest divergence points. This design places GEGRH between the speed of single-hypothesis planning and the conservatism of cross-hypothesis safety methods, with the stated aim of reducing oscillatory behavior while meeting a real-time planning budget (Damm et al., 30 Sep 2025).

1. Problem setting and motivation

The core problem addressed by GEGRH is navigation when a mobile ground robot lacks prior knowledge of its environment and must build a world model from sensor data. In such settings, consistent identification of obstacles and terrain features can be difficult due to noise and algorithmic shortcomings. A specific failure mode arises when regions of the cost map switch between being marked as obstacles and free space through successive planning cycles. The reported consequence is unstable behavior in which planners such as KEASL with A* or ARA* generate alternating, topologically distinct trajectories over time (Damm et al., 30 Sep 2025).

GEGRH is introduced against four comparison strategies. SH, or Single Hypothesis, uses only the most recent map and is therefore the fastest, but it oscillates when obstacle labels flip. VEH, or Valid in Every Hypothesis, requires every expanded edge to be collision-free in all maps in the stored history; this gives cross-hypothesis safety guarantees but is described as overly conservative and slower as the number of hypotheses grows. PEH, or Per-Edge Hypothesis, allows expansions valid in any hypothesis and invokes a sub-search for every conflicting edge, but this per-edge rerouting is computationally prohibitive under a 1-second field constraint. GEH, or Goal-Edge Hypothesis, defers rerouting to the point at which an edge reaches the goal, which is much faster than PEH but tends to concentrate expansions near the goal. GEGRH augments GEH with graph revision so that subsequent expansions focus on divergence points rather than remaining goal-centric (Damm et al., 30 Sep 2025).

A plausible implication is that GEGRH should be understood as a lazy multi-hypothesis kinodynamic planner: lazy because costly rerouting is postponed until a goal-edge is found, and multi-hypothesis because prior world models influence costs and search behavior even though only the primary hypothesis constrains final path validity.

2. Formal model and notation

The method operates on a sequence of world models

$W = \{M_1, M_2, \dots, M_T\},$

where $M_T$ is the most recent, or primary, hypothesis, and $M_1,\dots,M_{T-1}$ are prior hypotheses used to influence costs and risk-aware behavior. For each hypothesis $i$, the obstacle set is denoted by $O_i \subset \mathbb{R}^2$ or by the corresponding obstacle set in the robot’s configuration space (Damm et al., 30 Sep 2025).

The search graph is

$G = (V,E),$

built over a kinodynamic state lattice in KEASL. Nodes represent robot states, including pose and velocity state according to KEASL, and edges are trajectory segments generated by motion primitives that satisfy kinodynamic constraints. A divergence point is defined as the earliest node on a candidate path where per-hypothesis edge histories disagree, either because an edge is invalid in some $M_i$ due to collision with $O_i$ or because expansion histories differ between hypotheses. Divergence detection is supported by storing a per-hypothesis expansion history at each node (Damm et al., 30 Sep 2025).

Dynamic feasibility is expressed through a continuous-time state-control model

$\dot{x} = f(x,u),$

with $x(t)\in X$ and $M_T$0, together with feasibility bounds

$M_T$1

and constraints

$M_T$2

Each edge $M_T$3 corresponds to a finite-duration trajectory segment $M_T$4 generated by a motion primitive and collision-checked in one or more hypotheses (Damm et al., 30 Sep 2025).

VEH provides the strongest cross-hypothesis safety condition. Given a trajectory $M_T$5 and controls $M_T$6 over $M_T$7, VEH requires, for all $M_T$8,

$M_T$9

together with the same dynamics and feasibility constraints. GEGRH relaxes this requirement: expansions are allowed if valid in any hypothesis, but the final solution must be valid in the primary hypothesis $M_1,\dots,M_{T-1}$0 (Damm et al., 30 Sep 2025).

3. Cost structure and hypothesis-aware penalties

The base cost in the reported implementation is time-like path duration. The planner uses standard A*/ARA* scoring

$M_1,\dots,M_{T-1}$1

with heuristic inflation and with ARA* using $M_1,\dots,M_{T-1}$2 in the experiments (Damm et al., 30 Sep 2025). The accumulated path cost $M_1,\dots,M_{T-1}$3 sums edge costs based on traversal time or a length/time surrogate.

In PEH, GEH, and GEGRH, an expansion is permitted if it is valid in any hypothesis, but cost is adjusted to account for hypotheses in which the edge or path is invalid. In those cases, a Rerouter computes a feasible alternative trajectory segment from the divergence point to the goal. Costs across hypotheses are then combined by averaging with effectively uniform weights. The representative objective is written as

$M_1,\dots,M_{T-1}$4

where $M_1,\dots,M_{T-1}$5 is the additional cost incurred by hypothesis $M_1,\dots,M_{T-1}$6, for example a reroute cost when $M_1,\dots,M_{T-1}$7 is invalid under $M_1,\dots,M_{T-1}$8 (Damm et al., 30 Sep 2025).

At a goal-edge expansion, GEGRH performs a specific update. Let $M_1,\dots,M_{T-1}$9 be the last expansion to the goal found valid in at least one hypothesis. For each hypothesis $i$0, if the path is invalid in $i$1, the planner reroutes from the earliest divergence node $i$2 to the goal and computes a per-hypothesis cost $i$3; otherwise it uses the original primary goal-edge cost. The goal-edge cost is then updated by averaging:

$i$4

This averaging makes paths that cross inconsistent regions more expensive. The paper states that this nudges the planner toward solutions skirting uncertain obstacles or avoiding them when reroute penalties are large (Damm et al., 30 Sep 2025).

This suggests that GEGRH does not encode uncertainty through probabilistic occupancy or chance constraints. Instead, it converts historical inconsistency into deterministic cost inflation mediated by explicit alternate reroutes.

4. Algorithmic procedure

GEGRH differs from PEH and GEH in when it invokes rerouting and what it does afterward. PEH performs rerouting during every conflicting edge expansion. GEH postpones rerouting until a goal node is reached, updates the goal-edge cost, and reinserts the goal node into OPEN. GEGRH uses the same deferred rerouting trigger as GEH but then revises the graph so that the search resumes from the divergence structure rather than repeatedly working at the goal (Damm et al., 30 Sep 2025).

The GEGRH pseudocode reported in the paper is:

$O_i$2

The graph revision routine is:

$O_i$3

The stated purpose of graph revision is to shift expansion focus from the goal back to the earliest divergence point or points. Compared with GEH, this reduces repeated goal-centric work. Compared with PEH, it reduces the number of sub-search invocations because rerouting occurs only at goal expansions rather than at every conflicting edge (Damm et al., 30 Sep 2025).

A plausible interpretation is that graph revision functions as a targeted rollback mechanism. By pruning descendants of the divergence node, reorienting backpointers, and reopening the goal-expanded node, it converts a completed but hypothesis-inconsistent suffix into a revised search frontier anchored at the earliest problematic location.

5. Guarantees, complexity, and operational behavior

The principal guarantee stated for GEGRH is that the final solution must be valid in the primary hypothesis $i$5. This is enforced when extracting the final path. The method is more conservative than SH because it penalizes edges that would be invalid in prior hypotheses, and it is less conservative than VEH because it allows expansions that are invalid in some hypotheses while accounting for reroute costs (Damm et al., 30 Sep 2025).

No claim of global optimality across hypotheses is made. The paper states that ARA* provides bounded-suboptimality on the primary cost space when allowed to run to convergence with inflation deflation, but under 1-second time limits and hypothesis-weighted costs, practical suboptimality is expected. Likewise, while standard A*/ARA* completeness properties apply in static graphs with admissible heuristics and no time cap, GEGRH is time-bounded and modifies both costs and graph structure through revision, so practical completeness is not guaranteed within the time budget (Damm et al., 30 Sep 2025).

For computational complexity, the baseline A*/ARA* runtime is stated as $i$6 for typical implementations. GEGRH adds two main sources of overhead: occasional Rerouter sub-search from divergence point to goal, with cost proportional to subgraph size near divergence, and one graph revision per goal-edge update. The revision prunes descendants of the divergence node and reorients backpointers, focusing expansion where it matters. Empirically, GEGRH is reported to lower planning time relative to GEH by avoiding repeated work at the goal and to reduce sub-search frequency dramatically relative to PEH (Damm et al., 30 Sep 2025).

The practical behavior attributed to the method is therefore a three-way compromise. Against SH, it trades some speed for reduced oscillation. Against VEH, it trades strict cross-hypothesis safety for lower path cost and lower average runtime. Against PEH and GEH, it preserves the lazy evaluation strategy but removes a specific inefficiency associated with goal-centric expansion (Damm et al., 30 Sep 2025).

6. Empirical evaluation and comparative performance

The reported field evaluation used a Clearpath Robotics Warthog UGV in unstructured off-road terrain with online cost maps that often disagreed due to sensor noise, occlusions, and small state-estimation shifts. The planner was a KEASL lattice with ARA* using initial heuristic inflation $i$7 and a 1-second time limit per planning cycle. The dataset consisted of 221 planning problems, evaluated with two hypotheses ($i$8) and three hypotheses ($i$9). The full comparison focused on SH, VEH, and GEGRH, because PEH and GEH were omitted from full evaluation due to impractical runtimes and because GEGRH superseded GEH in practice (Damm et al., 30 Sep 2025).

The evaluation metrics were planning time and trajectory path duration in seconds.

Setting	Method	Result
Two hypotheses ($O_i \subset \mathbb{R}^2$0)	SH	planning $O_i \subset \mathbb{R}^2$1 s; path $O_i \subset \mathbb{R}^2$2 s
Two hypotheses ($O_i \subset \mathbb{R}^2$3)	VEH	planning $O_i \subset \mathbb{R}^2$4 s; path $O_i \subset \mathbb{R}^2$5 s
Two hypotheses ($O_i \subset \mathbb{R}^2$6)	GEGRH	planning $O_i \subset \mathbb{R}^2$7 s; path $O_i \subset \mathbb{R}^2$8 s
Three hypotheses ($O_i \subset \mathbb{R}^2$9)	SH	planning $G = (V,E),$0 s; path $G = (V,E),$1 s
Three hypotheses ($G = (V,E),$2)	VEH	planning $G = (V,E),$3 s; path $G = (V,E),$4 s
Three hypotheses ($G = (V,E),$5)	GEGRH	planning $G = (V,E),$6 s; path $G = (V,E),$7 s

For $G = (V,E),$8, GEGRH versus VEH yielded a $G = (V,E),$9 change in planning time, corresponding to $M_i$0 s, and a $M_i$1 change in path duration, corresponding to $M_i$2 s. For $M_i$3, GEGRH versus VEH yielded a $M_i$4 change in planning time, corresponding to $M_i$5 s, and a $M_i$6 change in path duration, corresponding to $M_i$7 s (Damm et al., 30 Sep 2025).

The key findings reported are that GEGRH finds lower-cost trajectories and faster average planning times than VEH, and that compared to SH it generates more conservative plans, reflected in longer paths, with only a small increase in average planning time. The paper further reports that PEH and GEH found more optimistic solutions than VEH but could not reliably meet the less-than-1-second field requirement, whereas GEGRH was the practical compromise (Damm et al., 30 Sep 2025).

The case studies described in the paper are not uniformly favorable. GEGRH can exploit narrow choke points that VEH rejects due to prior obstacles, thereby yielding shorter paths. However, in some maps, Rerouter overhead combined with the time limit can produce suboptimal final paths relative to VEH’s forced traversal through open areas. This prevents a simplified reading in which GEGRH dominates VEH in every instance; the reported advantage is empirical and average-case rather than absolute (Damm et al., 30 Sep 2025).

7. Implementation guidance, limitations, and interpretation

The implementation described in the paper uses a KEASL kinodynamic lattice with motion primitives and ARA* as the search engine. Nodes are augmented to store per-hypothesis expansion history, enabling detection of earliest divergence points and hypothesis-specific rerouting. During expansion, feasible edges are generated with motion primitives, collisions and velocity constraints are checked, and in GEGRH an edge is accepted if valid in any hypothesis, subject to the requirement that the final solution be collision-free in the primary hypothesis (Damm et al., 30 Sep 2025).

Several practical recommendations are stated. Hypothesis count $M_i$8 to $M_i$9 is described as practical; increasing the number of hypotheses raises both conservatism and runtime. The paper recommends maintaining per-hypothesis expansion histories in nodes, ensuring that the Rerouter reuses KEASL primitives and constraints, and applying graph revision after each goal-edge update to prevent goal-centric expansion loops and focus the search around divergence points. Uniform averaging is used in the present implementation, but the text notes that down-weighting older or lower-confidence hypotheses could reduce over-conservatism (Damm et al., 30 Sep 2025).

The listed limitations define the operational envelope of the method. GEGRH may still be conservative because averaging costs across hypotheses penalizes paths through inconsistent areas even when the latest map suggests those areas are free. In cluttered maps with many divergence points, Rerouter overhead and graph revision can consume the 1-second budget and yield suboptimal solutions. The method is also sensitive to hypothesis quality: if prior maps are noisy or misaligned, averaging may overweight spurious obstacles and steer the planner away from good primary-hypothesis routes. As the number of hypotheses grows, both conservatism and computational load increase, with the paper noting widening gaps in planning time and path duration from $O_i$0 to $O_i$1 (Damm et al., 30 Sep 2025).

Within the comparison set presented in the paper, GEGRH is positioned as the default choice for inconsistent environments under real-time constraints. SH is suitable when perception is stable and oscillations are rare or acceptable; VEH is appropriate when strict safety across map histories is required and conservatism is acceptable; GEH is conceptually simpler but is superseded by GEGRH because of goal-centric work concentration; and PEH is the highest-fidelity alternative but computationally prohibitive under tight time limits (Damm et al., 30 Sep 2025).

Taken together, these points characterize GEGRH as a kinodynamic multi-hypothesis planning strategy that operationalizes historical inconsistency as a cost signal rather than a hard feasibility constraint. Its central design choice is to postpone expensive cross-hypothesis reasoning until the planner has evidence of a candidate goal-reaching path, then restructure the graph to revisit only the earliest locations at which world-model disagreement matters.