PaperlayerView exact revision on arXiv

Source-linked AI summary

VikPath: A Vision Kansformer Framework for Effective Obstacle Avoidance in Self-Supervised Pathfinding

Junyao Wang, Yulin Xu, Mohammad Abdullah Al Faruque

arXiv:2608.22675v1cs.RO

TL;DR

Complex obstacle-rich environments expose limits in handcrafted pathfinding heuristics, while supervised learned planners depend on generated or annotated path labels. VikPath learns obstacle-aware representations self-supervisedly and integrates obstacle proximity and turning penalties into weighted A* search. Experiments report greater obstacle clearance, lower inference latency, smooth path generation, and favorable generalization to unseen environments.

  • Problem

    Handcrafted heuristics and supervised pathfinding labels provide limited guidance for complex environments, while existing methods often underemphasize obstacle clearance and trajectory smoothness.

  • Method

    VikPath learns an obstacle probability map with Vision Kansformer and incorporates obstacle proximity and turning penalties directly into smoothness-aware weighted A* search.

  • Results

    VikPath consistently improves obstacle-related path metrics and inference latency over evaluated baselines, with 3.28% greater obstacle clearance and 87.07% lower inference latency reported on average.

  • Takeaways & Limitations

    VikPath provides a pathfinding framework that balances obstacle clearance, trajectory smoothness, and search efficiency, with reported improvements and strong generalization to unseen environments.

Abstract

from arXiv · show

Pathfinding is a fundamental problem in artificial intelligence and autonomous systems. Traditional heuristic-based algorithms, such as A*, rely on predefined heuristic functions to guide the search process. Although effective in structured environments, their search efficiency can degrade substantially in complex, obstacle-rich scenarios, where handcrafted heuristics may provide limited guidance. Recent studies have explored learning-based approaches to improve pathfinding efficiency; however, most existing methods rely on supervised learning and require labels generated by conventional planners or obtained through manual annotation. As a result, their performance is inherently influenced by the quality of the underlying supervision and may degrade when the labeling heuristics fail to capture complex environmental structures. Moreover, existing methods primarily optimize for path length while paying limited attention to obstacle clearance and trajectory smoothness, which can lead to paths that are difficult or unsafe to execute in real-world environments. To address these limitations, we propose $\Design$, a self-supervised pathfinding framework that jointly considers obstacle proximity and path smoothness. At its core, our novel \textit{Vision Kansformer} module learns representations of obstacle distributions without relying on labeled trajectories, enabling the model to better adapt to complex environments. We further introduce a sharp-turn penalty to encourage smoother and more practically executable paths. Extensive experiments demonstrate that, compared with state-of-the-art (SOTA) approaches, $\Design$ achieves an average of 3.28\% greater obstacle clearance and 87.07\% lower inference latency while maintaining smooth path generation.

1 Introduction

Existing pathfinding methods can struggle with complex obstacle layouts because handcrafted heuristics and supervised labels provide limited environmental guidance. VikPath addresses these limitations through self-supervised obstacle representation learning and search that jointly considers obstacle proximity and path smoothness.

  • Motivation: Traditional heuristic-based pathfinding may inadequately capture complex obstacle configurations and global environmental patterns.A* commonly relies on handcrafted estimates and representations that provide limited spatial descriptions.
  • Motivation: Learning-based pathfinding methods often depend on expert or manually annotated trajectories, making them sensitive to supervisory-signal quality and labeling biases.Planner-generated labels can also limit adaptation across environments with diverse obstacle configurations.
  • VikPath: VikPath constructs probability maps that represent obstacle distributions, encouraging paths to avoid areas with high obstacle density.The representation provides explicit obstacle-density awareness during planning.
  • VikPath: VikPath is a self-supervised framework that learns environmental representations without heuristic-generated or manually annotated path labels.Its Vision Kansformer learns spatial structures directly from unlabeled environment maps.
  • Reported contributions: 87.07% lower inference latency and 3.28% greater obstacle clearance are reported for VikPath on average versus SOTA methods.The framework also incorporates obstacle awareness and path smoothness directly into planning.

2 Related Works

Related work spans classical graph-search algorithms, self-supervised visual representation learning, and flexible neural function approximators. These foundations motivate learning richer spatial structure for pathfinding beyond conventional geometric heuristics.

  • Classical pathfinding: Dijkstra’s algorithm uses accumulated path cost, while A* additionally uses a heuristic estimate of remaining cost to the goal.Common Euclidean and Manhattan heuristics encode limited information about obstacle configurations.
  • Self-supervised representation learning: Masked image modeling learns visual representations by reconstructing masked portions of inputs from visible context without manual annotations.This objective encourages models to capture contextual relationships and structural information.
  • Kolmogorov-Arnold Networks: Kolmogorov-Arnold Networks replace fixed MLP activations and scalar weights with learnable univariate functions for flexible function approximation.These functions are typically parameterized using B-splines.

3.1 Problem Formulation

VikPath formulates pathfinding as a two-stage process: learning an obstacle-aware probability map from an environment map, then performing obstacle- and smoothness-aware weighted A* search. The formulation represents environments as binary occupancy maps while enriching free-space locations with learned spatial context.

  • Problem formulation: VikPath takes an environment map, start location, and goal location and seeks a collision-free path balancing efficiency, obstacle clearance, and trajectory smoothness.Its two stages are representation learning and weighted A* path search.
  • Environment representation: The environment is represented as a binary occupancy map I ∈ {0, 1}^m×n, distinguishing free locations from obstacles.The formulation separately defines start and goal locations and restricts feasible paths to adjacent free-space locations.
  • Learned representation: Because binary occupancy does not encode surrounding obstacle context, Vision Kansformer constructs an obstacle probability map from the environment.This map represents spatial information such as nearby-obstacle proximity and obstacle distributions.
  • Learned representation: Each P_i,j is a normalized obstacle-related score, with larger values indicating stronger obstacle-related influence and smaller values indicating relatively open regions.The map is incorporated into obstacle- and smoothness-aware weighted A* search.
  • Path search: The cumulative search cost jointly accounts for movement distance, obstacle proximity, and changes in path direction, while a weighted heuristic provides goal-directed guidance.This formulation defines search states using arrival direction because turning cost depends on how a location is reached.

3.2 Learning-based Obstacle Probability Map

VikPath learns obstacle-aware spatial representations from unlabeled occupancy maps through masked reconstruction and a Vision Kansformer encoder. Complementary objectives and a reconstruction head produce a dense probability map that guides planning away from obstacle-influenced regions.

  • Masking and representation learning: Masked reconstruction encourages VikPath to infer obstacle structures from visible spatial context rather than relying only on individual occupancy values.The input occupancy map is masked before reconstruction, promoting contextual spatial dependency learning.
  • Vision Kansformer: Vision Kansformer combines self-attention for long-range spatial dependencies with KAN transformations for flexible nonlinear refinement.Patch embeddings receive positional information, and each block replaces the conventional feed-forward network with a KAN layer and residual connection.
  • Training objectives: The training objective combines contrastive, reconstruction, and perceptual losses to learn discriminative representations, obstacle-proximity-aware structures, and higher-level spatial information.The three losses are weighted by β, λ, and α, respectively.
  • Training objectives: Obstacle-proximity-aware reconstruction targets assign larger values to obstacles and progressively smaller values to free-space regions farther from obstacle boundaries.Masked-region reconstruction encourages the model to distinguish open regions from areas close to obstacles.
  • Probability map generation: The reconstruction head converts final representations into a normalized dense obstacle probability map used to guide path search away from strong obstacle influence.The map is formed by refining spatial features, aggregating channels, and applying min–max normalization; larger values indicate stronger obstacle-related influence.

3.3 Obstacle- and Smoothness-Aware Weighted A* Pathfinding

VikPath performs weighted A* search using states that retain predecessor information, allowing traversal costs to account jointly for movement, obstacle proximity, and turning smoothness. These costs influence search directly rather than being applied only during post-processing.

  • State representation: Each search state stores the current location and its predecessor, distinguishing arrivals with different subsequent turning costs.The initial state is X_S = (∅, S), and successor states are formed from the current location and neighboring location.
  • Weighted heuristic: The weighted evaluation function combines cumulative path cost with a heuristic estimate of remaining travel cost.The heuristic weight w_h is at least 1; w_h = 1 recovers standard A* weighting, while larger values emphasize heuristic guidance.
  • Traversal cost: Each transition cost combines movement distance, an obstacle-related penalty, and a smoothness penalty for direction changes.The obstacle score P(Q_n) and weights w_o and w_s control the contributions of obstacle proximity and smoothness.
  • Search procedure: The search expands the lowest-priority state, updates improved successor costs and parents, and reopens expanded states when lower-cost paths are found.When the goal is reached, the final path is recovered by backtracking through stored parent states.
  • Search objectives: Turning penalties are incorporated into transition costs, so obstacle clearance and path smoothness influence search rather than only post-search smoothing.Continuing in the same direction has zero turning cost, while larger direction changes receive larger penalties.

4 RESULT

The experiments evaluate VikPath on public pathfinding datasets against representative search-based and learning-based methods using safety, clearance, smoothness, cost, exploration, and latency measures. The section defines the datasets, implementation setting, comparison methods, and evaluation metrics.

  • Datasets: VikPath is evaluated on three public pathfinding datasets under settings commonly used in prior learning-based pathfinding studies.The datasets include diverse obstacle configurations and grid-world environments, with augmentation applied to the maps.
  • Implementation: Training randomly masks 15%–20% of each occupancy map and reconstructs the masked regions using the stated objectives.The Vision Kansformer encoder uses 12 layers, 6 attention heads, and an embedding dimension of 384.
  • Baselines: The comparisons include Weighted A* Search, Improved JPS, and other representative search-based and learning-based methods.Weighted A* emphasizes heuristic guidance, while Improved JPS adds path-safety inflation near obstacles.
  • Evaluation metrics: Evaluation covers path safety, obstacle clearance, trajectory smoothness, obstacle-related traversal cost, search efficiency, and inference latency.The metrics include path safety score, average obstacle distance, angular deviation score, average obstacle-related cost, node exploration reduction, and inference latency.
  • Metric interpretation: Higher clearance and safety scores indicate safer paths, while lower angular deviation, obstacle-related cost, and latency indicate more favorable outcomes.Node exploration reduction is higher when fewer states are explored relative to standard A*.

4.5 Qualitative results

Figure 3 visualizes obstacles, endpoints, generated paths, and VikPath’s obstacle probability map. The qualitative examples emphasize how the learned map and smoothness penalty guide paths toward safer, clearer, and less abrupt trajectories.

  • Probability map: In the probability map, lighter gray regions indicate stronger obstacle-related scores, darker gray regions indicate greater clearance, and black regions represent relatively open free space.White obstacle regions are non-traversable.
  • Figure encoding: Figure 3 marks obstacles in white, starting points in cyan, destinations in yellow, and generated paths in red.The rightmost column shows VikPath’s probability map.
  • Qualitative comparison: Compared with baseline methods, VikPath more effectively avoids strongly obstacle-influenced regions and produces paths with greater clearance from nearby obstacles.The learned probability map provides global spatial information before and during path search.
  • Trajectory behavior: The explicit smoothness penalty discourages abrupt direction changes while balancing obstacle clearance, path efficiency, and trajectory smoothness.

4.6 Quantitative results

VikPath improves obstacle clearance, safety, search efficiency, and inference latency across baseline and unseen-environment evaluations, while exposing a trade-off between smoothness and clearance.

  • Baseline comparison: VikPath achieves higher path safety and average obstacle distance than baseline methods, indicating greater obstacle clearance.Table 1 reports consistently stronger safety-related metrics for VikPath.
  • Baseline comparison: 3.28% higher path safety score, 5.62% higher average obstacle distance, and 4.09% lower average obstacle-related cost distinguish VikPath from the second-best method.These averages summarize the reported Table 1 comparison.
  • Efficiency: 87.06% lower inference latency than the second-best baseline makes VikPath the fastest method across all three datasets.VikPath also ranks first in node exploration reduction on MP and second on BF and TMP.
  • Generalization: On unseen City and Game datasets without fine-tuning, VikPath maintains favorable safety, obstacle-distance, cost, exploration, and latency results.The model transfers across obstacle layouts and spatial structures that differ from training data.

4.7 Ablation Study

Ablations show that learned obstacle-aware cost and explicit smoothness penalties complement each other, producing a favorable balance among clearance, smoothness, and search efficiency.

  • Cost-function ablation: Jointly incorporating learned obstacle-related cost and smoothness penalty provides the most favorable overall pathfinding performance.The ablation compares full VikPath with improved A* and variants removing either penalty.
  • Cost-function ablation: 0.67% higher path safety, 6.83% higher average obstacle distance, and 43.00% lower average obstacle-related cost are achieved against improved A*.The same comparison reports 26.90% lower angular deviation and 74.25% higher node exploration reduction.
  • Component roles: The learned obstacle probability map supplies environment-dependent guidance, while the smoothness penalty discourages abrupt turns during search.Their roles are complementary rather than interchangeable.
  • Representation ablation: Vision Kansformer is evaluated against SiT and Vision Mamba using MSE, PSNR, and SSIM for masked-reconstruction quality.Lower MSE and higher PSNR and SSIM indicate better reconstruction quality.

5 Conclusions

VikPath combines self-supervised obstacle-aware representations with smoothness-aware search to balance clearance, trajectory smoothness, and efficiency. Experiments report improved baselines, low latency, and generalization to unseen environments.

  • Conclusion: VikPath learns an obstacle probability map through Vision Kansformer and incorporates obstacle proximity and turning penalties directly into search.The framework is designed for complex obstacle-rich environments.
  • Conclusion: The reported outcomes include a favorable balance among obstacle clearance, trajectory smoothness, and search efficiency.The conclusion also reports low inference latency and strong generalization to unseen environments.
Loading 2608.22675v1…