CogniPlan: Uncertainty-Guided Path Planning

• CogniPlan is a path-planning framework for mobile robots in partially observed environments, integrating conditional layout prediction with graph-attention planning.
• It employs a conditional WGAN-GP trained generator to predict plausible completions of unknown maps, enabling uncertainty-aware navigation decisions.
• Empirical results demonstrate reduced travel lengths in autonomous exploration and point-goal navigation compared to conventional frontier-based and reactive methods.

Searching arXiv for the cited papers to ground the article and verify identifiers. {"query":"CogniPlan (Wang et al., 5 Aug 2025) seneca (Bohnen et al., 21 Apr 2026) PlanGlow (Chun et al., 16 Apr 2025) StatePlane (Annapureddy et al., 13 Mar 2026)", "max_results": 10} I’ll use the arXiv search tool to verify the core papers and identifiers before writing the article. {"query":"\"CogniPlan: Uncertainty-Guided Path Planning with Conditional Generative Layout Prediction\"","max_results":5} CogniPlan is a path-planning framework for mobile robots in unknown environments that couples a conditional generative layout predictor with a graph-attention planner trained with deep reinforcement learning. It is designed for two coupled tasks—autonomous exploration and point-goal navigation—under partial observability, where the robot incrementally builds an occupancy map yet must still estimate potential information gain and plan effectively through unseen space. Its defining idea is to predict multiple plausible completions of the unknown map, treat them as a form of cognitive map, and then plan on a graph constructed over predicted free space rather than relying only on observed frontiers or purely reactive policies (Wang et al., 5 Aug 2025).

1. Problem setting and defining characteristics

CogniPlan operates in environments where only part of the map is known at any moment and large regions remain unknown and could be free or obstructed. The framework addresses two tasks. In autonomous exploration, the objective is to reach high, ideally full, coverage of free space while minimizing travel distance or time. In point-goal navigation, the objective is to reach a given goal position with minimal path length even when obstacles near the goal have not yet been observed (Wang et al., 5 Aug 2025).

The environment is represented as a 2D occupancy grid. The paper writes the environment as

$\mathcal{E} = \mathcal{E}_f \cup \mathcal{E}_o,$

and the belief map as

$$\mathcal{M} = \mathcal{M}_u \cup \mathcal{M}_f \cup \mathcal{M}_o,$$

where $\mathcal{M}_u$ denotes unknown cells, $\mathcal{M}_f$ known free cells, and $\mathcal{M}_o$ known occupied cells. A trajectory is a sequence of waypoints in free space,

$\psi : \{0,1,2,\dots\} \to \mathcal{E}_f.$

The framework is motivated by a specific failure mode of conventional methods. Frontier-based planners are described as mostly greedy and purely map-based, whereas many DRL planners are described as purely reactive and as not explicitly imagining unobserved layouts. CogniPlan addresses this by using multiple plausible predicted layouts to guide planning under uncertainty, explicitly coupling exploration and exploitation rather than treating unknown space as a blank region to be greedily expanded into (Wang et al., 5 Aug 2025).

2. Conditional generative layout prediction

The generative component takes a partially observed occupancy grid $\mathcal{M}$ and a layout conditioning vector $z$, then predicts a complete layout $\hat{\mathcal{M}}$ that fills unknown regions with plausible free or occupied structure. Training uses procedurally generated maps of three layout types—room-like, tunnel, and outdoor—and the generator is a lightweight convolutional network with approximately $0.35$M parameters. Inputs include the occupancy map belief, a binary mask of unknown cells, and the conditioning vector $$\mathcal{M} = \mathcal{M}_u \cup \mathcal{M}_f \cup \mathcal{M}_o,$$0 broadcast and concatenated as channels (Wang et al., 5 Aug 2025).

The generator is trained with a conditional WGAN-GP objective. The paper combines adversarial and reconstruction terms, including an $$\mathcal{M} = \mathcal{M}_u \cup \mathcal{M}_f \cup \mathcal{M}_o,$$1 term, a spatially discounted mask term that emphasizes regions near known cells, and an F1 term. Training includes a $$\mathcal{M} = \mathcal{M}_u \cup \mathcal{M}_f \cup \mathcal{M}_o,$$2k-step warmup phase using only reconstruction terms, followed by $$\mathcal{M} = \mathcal{M}_u \cup \mathcal{M}_f \cup \mathcal{M}_o,$$3k full WGAN-GP training iterations. This design places higher weight on regions near the observed map, which the paper identifies as more critical for robot safety and connectivity (Wang et al., 5 Aug 2025).

At inference time, the true layout type is unknown, so CogniPlan does not rely on a single conditioning vector. Instead it uses a set of conditioning vectors $$\mathcal{M} = \mathcal{M}_u \cup \mathcal{M}_f \cup \mathcal{M}_o,$$4, including one-hot vectors for room, tunnel, and outdoor, soft one-hot vectors, and a uniform vector. For each $$\mathcal{M} = \mathcal{M}_u \cup \mathcal{M}_f \cup \mathcal{M}_o,$$5, the generator produces

$$\mathcal{M} = \mathcal{M}_u \cup \mathcal{M}_f \cup \mathcal{M}_o,$$6

These predictions are averaged to obtain

$$\mathcal{M} = \mathcal{M}_u \cup \mathcal{M}_f \cup \mathcal{M}_o,$$7

which is interpreted cell-wise as the probability that each cell is free. The paper explicitly treats this as uncertainty-aware prediction: disagreement among the generated layouts corresponds to uncertainty about unseen structure, while agreement yields a more confident cognitive map (Wang et al., 5 Aug 2025).

3. Graph construction and graph-attention planning

CogniPlan does not plan directly on the occupancy grid. Instead it constructs a collision-free graph $$\mathcal{M} = \mathcal{M}_u \cup \mathcal{M}_f \cup \mathcal{M}_o,$$8 over predicted free space. Nodes are uniformly distributed over predicted free cells, and edges are added when distance is below a threshold and there is line-of-sight free space. In the reported setup, graph nodes are spaced $$\mathcal{M} = \mathcal{M}_u \cup \mathcal{M}_f \cup \mathcal{M}_o,$$9 px apart and edges are created within radius $\mathcal{M}_u$0 px, with at most $\mathcal{M}_u$1 neighbors (Wang et al., 5 Aug 2025).

Each node is augmented with a feature vector

$\mathcal{M}_u$2

where $\mathcal{M}_u$3 is the node coordinate, $\mathcal{M}_u$4 indicates whether the node lies in known space or predicted space, $\mathcal{M}_u$5 is the averaged free-space probability from $\mathcal{M}_u$6, $\mathcal{M}_u$7 is a utility derived from the number of frontiers observable from the node, and $\mathcal{M}_u$8 is a guidepost flag indicating that the node lies on a candidate shortest path to the nearest frontier in known regions. For point-goal navigation, an additional direction vector from the node to the goal is concatenated (Wang et al., 5 Aug 2025).

The planner itself is a graph-attention policy network with an encoder-decoder structure. A feed-forward layer first maps node features into embeddings, after which $\mathcal{M}_u$9 layers of masked self-attention aggregate information over graph neighborhoods. The robot’s current node embedding is then used as a query in the decoder, and a final single-head attention over neighboring nodes yields attention scores interpreted as a discrete action policy over waypoint choices. The critic uses the same graph-attention structure but receives privileged information from the ground-truth graph $\mathcal{M}_f$0, a form of privileged learning intended to improve value estimation during training (Wang et al., 5 Aug 2025).

Training uses Soft Actor-Critic in a discrete action space. For exploration, the reward combines positive reward for increased coverage or discovering new frontiers, a large positive terminal reward when exploration is complete, and a penalty proportional to travel distance. For navigation, the reward includes a positive terminal reward at the goal, a dense shaping reward for reducing distance to the goal, and a penalty for path length (Wang et al., 5 Aug 2025).

4. Empirical performance and ablation results

The framework is evaluated on two datasets described as “hundreds of maps and realistic floor plans,” and also in a high-fidelity simulator and on hardware. On $\mathcal{M}_f$1 unseen maps for exploration, mean travel length to full exploration is reported as $\mathcal{M}_f$2 for CogniPlan, versus $\mathcal{M}_f$3 for ARiADNE+, $\mathcal{M}_f$4 for TARE Local, $\mathcal{M}_f$5 for NBVP, $\mathcal{M}_f$6 for Utility frontier, $\mathcal{M}_f$7 for Nearest frontier, and $\mathcal{M}_f$8 for Inpaint+TARE. The paper states that ARiADNE+ is $\mathcal{M}_f$9 worse than CogniPlan, while Inpaint+TARE is $\mathcal{M}_o$0 worse (Wang et al., 5 Aug 2025).

On $\mathcal{M}_o$1 unseen maps for point-goal navigation, mean path length is reported as $\mathcal{M}_o$2 for CogniPlan, compared with $\mathcal{M}_o$3 for D*Lite, $\mathcal{M}_o$4 for Inpaint+A*, $\mathcal{M}_o$5 for CA, and $\mathcal{M}_o$6 for BIT. The gap to D*Lite is $\mathcal{M}_o$7, while CA is $\mathcal{M}_o$8 worse and Inpaint+A is $\mathcal{M}_o$9 worse (Wang et al., 5 Aug 2025).

Task	Method	Result
Exploration	CogniPlan	306.4 mean travel length
Exploration	ARiADNE+	329.3 mean travel length
Exploration	Inpaint+TARE	372.3 mean travel length
Navigation	CogniPlan	222.1 mean path length
Navigation	D*Lite	224.6 mean path length
Navigation	CA	253.9 mean path length

A central ablation concerns the number of layout predictions. Using $\psi : \{0,1,2,\dots\} \to \mathcal{E}_f.$0, the paper reports that more predictions improve performance for both exploration and navigation, with clear gains from $\psi : \{0,1,2,\dots\} \to \mathcal{E}_f.$1 to $\psi : \{0,1,2,\dots\} \to \mathcal{E}_f.$2 and smaller incremental gains from $\psi : \{0,1,2,\dots\} \to \mathcal{E}_f.$3 to $\psi : \{0,1,2,\dots\} \to \mathcal{E}_f.$4. Another key ablation isolates the planner from the predictor. The classical pipelines Inpaint+TARE and Inpaint+A* show that naive use of predicted layouts is not enough; indeed, Inpaint+TARE performs worse than TARE alone, especially early in exploration, which the paper attributes to noisy predictions causing unstructured zig-zag behavior. This directly rebuts the misconception that a generative predictor can simply be inserted upstream of a conventional planner without uncertainty-aware integration (Wang et al., 5 Aug 2025).

The simulation and hardware results extend the same pattern. In a medium-scale indoor Gazebo environment of approximately $\psi : \{0,1,2,\dots\} \to \mathcal{E}_f.$5 m $\psi : \{0,1,2,\dots\} \to \mathcal{E}_f.$6 $\psi : \{0,1,2,\dots\} \to \mathcal{E}_f.$7 m, total travel distance is $\psi : \{0,1,2,\dots\} \to \mathcal{E}_f.$8 m for CogniPlan, versus $\psi : \{0,1,2,\dots\} \to \mathcal{E}_f.$9 for HPHS, $\mathcal{M}$0 for TARE, $\mathcal{M}$1 for frontier-based exploration, and $\mathcal{M}$2 for TDLE. In a large-scale indoor environment of approximately $\mathcal{M}$3 m $\mathcal{M}$4 $\mathcal{M}$5 m, CogniPlan achieves $\mathcal{M}$6 m total travel distance, compared with $\mathcal{M}$7 for TARE and $\mathcal{M}$8 for DSVP. On a real robot in a $\mathcal{M}$9 m $z$0 $z$1 m cluttered lab, the paper reports roughly $z$2 min to almost fully explore the lab and a path length of approximately $z$3 m using the same trained models as in simulation (Wang et al., 5 Aug 2025).

5. Broader conceptual interpretations

The concrete system named CogniPlan is the robotic framework above, but the surrounding literature in the supplied corpus treats “CogniPlan” as a useful label for a broader family of cognitively informed planners. This suggests a conceptual extension rather than an additional definition. In that broader sense, several papers emphasize three recurrent themes: metacognitive control of planning effort, persistent structured state, and self-reflective adjustment of plans.

The metacognitive theme is explicit in “Have I done enough planning or should I plan more?” (He et al., 2022), which models planning as a meta-level MDP over computations with a termination action and reports that the REINFORCE family explains approximately $z$4 of participants at the family level, with “REINFORCE + PR” the best individual model. “The Efficiency of Human Cognition Reflects Planned Information Processing” (Ho et al., 2020) formalizes “planning one’s plans” with a recursive Bellman objective that trades task reward against information-theoretic planning cost. Together, these papers suggest that a cognitively informed CogniPlan can be interpreted not only as a planner over actions but also as a controller over how much planning to do.

The persistent-state theme appears in “StatePlane: A Cognitive State Plane for Long-Horizon AI Systems Under Bounded Context” (Annapureddy et al., 13 Mar 2026), which defines a model-agnostic cognitive state plane over episodic, semantic, and procedural state, with bounded reconstruction

$z$5

goal-conditioned retrieval, and adaptive forgetting. A related conversational planning interpretation appears in “seneca: A Personalized Conversational Planner” (Bohnen et al., 21 Apr 2026), which is described as consisting of “a database, a user interface combining a conversational agent with a structured work item view, and a processor that moves information between them.” These works do not redefine the robotic CogniPlan framework, but they do indicate how the name can extend to long-horizon and personalized planning systems that combine structured state with adaptive reasoning.

6. Limitations and future directions

The robotic CogniPlan paper explicitly identifies several limitations. First, the inpainting network expects a fixed-size grid such as $z$6, which assumes a bounded workspace and may limit generalization to much larger or differently scaled environments. Second, the coupling between predictor and planner is one-way: the generator is pre-trained and frozen, and gradients do not flow back from the planner. Third, each new prediction currently triggers a complete rebuild of the graph, creating computational overhead. Fourth, there is a domain gap in out-of-distribution environments such as KTH floor plans and large Gazebo scenes, where predicted layouts can differ noticeably from ground truth. Fifth, the current system focuses on single-robot 2D planning rather than multi-robot collaboration or full 3D exploration (Wang et al., 5 Aug 2025).

The same paper frames these as concrete directions for further work: resolution-agnostic or multi-scale predictors, end-to-end coupling between generator and planner, incremental graph updates, more diverse training maps, incorporation of semantics, and extension to multi-robot or 3D settings. A broader implication, suggested by the related conceptual literature, is that future CogniPlan variants may combine uncertainty-guided spatial planning with explicit memory, reflection, or latent planning mechanisms. That broader synthesis remains an interpretation rather than a claim of the original robotics system, but it captures why CogniPlan has become a useful reference point for cognitively informed planning across several research areas (Wang et al., 5 Aug 2025).