💻 computer science

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

VikPath is a self-supervised framework featuring a novel Vision Kansformer module that learns obstacle distributions without labeled data to generate smoother, safer, and more efficient paths with significantly lower inference latency than state-of-the-art methods.

Original authors: Junyao Wang, Yulin Xu, Mohammad Abdullah Al Faruque

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Junyao Wang, Yulin Xu, Mohammad Abdullah Al Faruque

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of robots and autonomous vehicles, the ability to find a way from point A to point B is a fundamental challenge. For decades, computers have solved this using mathematical rules that treat the world like a grid of squares, calculating the shortest distance while trying to avoid walls. These traditional methods work well in simple, predictable rooms, but they often struggle in complex, cluttered spaces where obstacles are packed tightly together. The old rules tend to produce paths that are technically the shortest but dangerously close to objects, or routes that require sudden, jerky turns that a real vehicle cannot physically make. Furthermore, many modern attempts to teach computers to navigate better have relied on showing them thousands of examples of perfect paths created by humans or other computers. This approach has a flaw: if the examples are imperfect or biased, the learning system inherits those same mistakes, and it often fails when faced with a new environment it has never seen before.

A team of researchers at the University of California, Irvine, has developed a new approach called VikPath that changes how machines learn to navigate without needing a teacher to show them the way. Instead of memorizing pre-made routes, their system learns to understand the layout of an environment on its own, much like a person might glance at a crowded room and instinctively know where the open spaces are and where the clusters of furniture are too tight to squeeze through. The researchers created a special learning module that looks at a map of obstacles and fills in the missing details, learning to predict where obstacles are likely to be and how close a safe path should stay to them. This allows the system to generate a "probability map," a mental guide that highlights safe, open corridors and warns against areas that are too crowded, even if those areas are technically empty space.

The core of this new system is a unique architecture that combines two powerful ideas to understand spatial relationships. It uses a method that hides parts of a map and forces the computer to guess what is missing based on the surrounding context, a technique that helps it learn the structure of the environment rather than just memorizing specific patterns. To make these guesses more accurate and flexible, the system uses a mathematical framework that can adapt its internal rules to fit the specific shapes and sizes of the obstacles it encounters. Once the system has learned to read the map and understand where the dangers lie, it uses a refined search process to plot a course. This process does not just look for the shortest line; it actively penalizes paths that get too close to obstacles and discourages sharp, sudden turns. The result is a route that is not only efficient but also smooth and safe, avoiding the tight squeezes and jerky movements that plague older methods.

When the researchers tested their system against the best existing methods, the results were striking. In a series of tests using three different types of complex environments, VikPath consistently found paths that were significantly safer than those found by traditional algorithms. On average, the paths generated by the new system kept a distance of 3.28 percent greater from obstacles compared to the next best method, meaning the robot would be less likely to scrape against a wall or get stuck in a narrow gap. At the same time, the paths were much smoother, with fewer abrupt changes in direction, making them easier for a physical vehicle to follow. Perhaps most surprisingly, the system was incredibly fast. It completed its calculations in a fraction of the time required by other advanced learning-based methods, showing an 87.07 percent reduction in the time it took to find a path. This speed comes from the fact that the system does not need to process complex, pre-labeled examples or perform extra steps to smooth out the route after it is found; the safety and smoothness are built into the search process from the very beginning.

The true test of any navigation system is whether it can handle a world it has never seen before. The researchers trained their system on one set of maps and then asked it to navigate entirely different environments, including layouts resembling city streets and video game levels, without any further training. The system adapted immediately, maintaining its ability to find safe, smooth paths and avoiding the need for a human to re-teach it the rules of the new environment. This suggests that the system has learned a deep understanding of how obstacles are distributed in space, rather than just memorizing specific patterns. By teaching machines to understand the shape of danger and the value of a smooth journey, this work offers a new way for autonomous systems to move through our complex, obstacle-filled world with greater confidence and efficiency.

Technical Summary: VikPath

Problem Statement

Pathfinding is a fundamental challenge in artificial intelligence and autonomous systems, requiring the determination of efficient, collision-free routes from a start to a destination. Traditional heuristic-based algorithms, such as A*, rely on predefined heuristics (e.g., Euclidean or Manhattan distance) and grid/graph representations. While effective in structured environments, these methods often struggle in complex, obstacle-rich scenarios where handcrafted heuristics fail to capture intricate spatial structures. Furthermore, classical planners typically optimize solely for path length and collision avoidance, often neglecting practical factors like obstacle clearance (proximity to obstacles) and trajectory smoothness. This can result in paths that pass dangerously close to obstacles or contain abrupt turns, rendering them suboptimal or unsafe for real-world execution.

Recent learning-based approaches have attempted to address these issues but often rely on supervised learning using labels generated by conventional planners or manual annotation. This dependency limits generalization, as the learned models inherit the biases and limitations of the underlying supervision (e.g., prioritizing shortest paths over safety). Additionally, existing methods often treat path generation and smoothing as separate stages (e.g., using post-processing Bézier curves), which can compromise obstacle clearance or require additional feasibility checks.

Methodology

The paper proposes VikPath, a self-supervised pathfinding framework designed to jointly optimize obstacle proximity, path smoothness, and computational efficiency without relying on expert-generated path labels. The framework operates in two stages:

1. Learning Obstacle-Aware Representations (Vision Kansformer)

The first stage learns a dense obstacle probability map directly from unlabeled environment maps using a novel Vision Kansformer module.

Architecture: The model combines the global context modeling of Self-Attention (from Vision Transformers) with the flexible nonlinear transformations of Kolmogorov-Arnold Networks (KANs). Unlike standard Transformers that use Feed-Forward Networks (FFNs), VikPath replaces these with KAN layers, which utilize learnable univariate functions (parameterized by B-splines) to refine representations.
Training Strategy: The module is trained using a masked modeling approach (inspired by Masked Image Modeling).
- Masking: Random patches of the binary occupancy map are masked.
- Objectives: The model is trained to reconstruct the masked regions using three complementary losses:
  1. Reconstruction Loss: The target is not the original binary map but an obstacle-proximity-aware target. This target is generated by iteratively dilating obstacle regions and assigning higher values to obstacles and their immediate surroundings, encouraging the model to learn spatial proximity.
  2. Contrastive Loss: A pre-trained Vision Transformer (ViT) acts as a teacher to provide reference representations, enhancing the discriminability of the learned features.
  3. Perceptual Loss: Ensures the preservation of high-level structural information during reconstruction.
Output: The reconstruction head generates a normalized probability map $P \in [0, 1]^{m \times n}$ , where higher values indicate regions with strong obstacle influence.

2. Obstacle- and Smoothness-Aware Path Search

The second stage performs pathfinding using a modified Weighted A (WA)** algorithm that incorporates the learned probability map and a smoothness penalty directly into the search cost.

State Definition: To account for smoothness, the search state is defined as a tuple $(Q_p, Q)$ , representing the current location and its predecessor, rather than just the current location.
Cost Function: The traversal cost $c(X, X_n)$ $c (X, X_{n})$ for a transition includes:
1. Movement Distance: Standard step cost.
2. Obstacle Proximity Penalty: A weighted term $w_o P(Q_n)$ derived from the learned probability map, penalizing paths through high-obstacle-density regions.
3. Sharp-Turn Penalty: A cost $c_{smooth}$ calculated based on the angle between incoming and outgoing motion vectors, discouraging abrupt direction changes.
Search: The algorithm uses a weighted heuristic to guide the search toward the goal while balancing the safety and smoothness costs, eliminating the need for separate post-processing smoothing steps.

Key Contributions

VikPath Framework: A self-supervised pathfinding system that learns environmental representations without heuristic-generated or manually annotated path labels, jointly incorporating obstacle awareness and path smoothness.
Vision Kansformer Module: A novel architecture integrating KANs into a Transformer structure with masked modeling. It captures both local and global spatial dependencies to construct fine-grained obstacle probability maps, achieving an average of 3.28% greater obstacle clearance compared to state-of-the-art (SOTA) methods.
Unified Search Objective: An enhancement of the Weighted A* algorithm that explicitly includes obstacle proximity and a sharp-turn penalty in the cost function. This enables the generation of safe, smooth paths without requiring a separate Bézier-curve-based post-processing stage.
Efficiency: The framework achieves an average of 87.07% lower inference latency than SOTA methods while maintaining high path quality.

Experimental Results

The authors evaluated VikPath on three public datasets (Motion Planning, Bugtrap Forest, and Tiled Motion Planning) and two unseen datasets (City and Game) to test generalization.

Performance Metrics: VikPath consistently outperformed baselines (including WA*, Improved JPS, SAIL-SL, BB-A*, FDW-A*, and APF-A*) in:
- Path Safety Score: Higher accumulated clearance from obstacles.
- Average Obstacle Distance: Greater mean distance to nearest obstacles.
- Smoothness: Lower angular deviation scores (indicating smoother trajectories) compared to methods that prioritize safety over smoothness.
- Search Efficiency: Significant reduction in node exploration compared to standard A* and other baselines.
- Latency: The lowest inference latency across all datasets.
Generalization: When applied to unseen datasets (City and Game) without fine-tuning, VikPath maintained superior performance in safety, clearance, and efficiency, demonstrating robust cross-dataset generalization.
Ablation Studies:
- Removing the learned obstacle penalty or the smoothness penalty resulted in significant drops in safety and smoothness metrics, confirming the necessity of both components.
- Comparing reconstruction techniques showed that the Vision Kansformer outperformed Vision Transformers and Vision Mamba in reconstructing obstacle structures (lower MSE, higher PSNR/SSIM).

Significance

The paper claims that VikPath addresses critical limitations in existing pathfinding approaches by eliminating reliance on expert-generated labels and unifying obstacle clearance with trajectory smoothness into a single, efficient search process. By leveraging self-supervised learning to understand obstacle distributions and integrating these insights directly into the pathfinding cost function, VikPath generates paths that are not only shorter but also safer and more executable in real-world environments. The framework's ability to achieve high performance with low inference latency makes it particularly suitable for latency-sensitive autonomous navigation applications in complex, obstacle-rich scenarios.