Robotic Hierarchical Graph Neurons. A novel implementation of HGN for swarm robotic behaviour control

Phillip Smith, Aldeida Aleti, Vincent C. S. Lee, Robert A. Hunjet, Asad KhanExpert Systems with Applications 2019 · DOI

Abstract

This paper introduces Robotic Hierarchical Graph Neurons (R-HGN), a decentralized pattern recognition framework designed for autonomous behavior switching in robotic swarms. By extending the Hierarchical Graph Neuron (HGN) architecture with dynamic neuron creation and probabilistic fusion, R-HGN enables swarms to identify environmental contexts from continuous sensor data and reach a consensus on optimal behaviors. Experimental results demonstrate high classification accuracy (92.4%–99.1%) and robust performance in unseen and hybrid environments, outperforming traditional static behavior selection methods.

Introduction

In swarm robotics, the ability to adapt to diverse environments is critical. A single behavior, such as wall-following or dispersion, is rarely optimal across all scenarios. While specific behaviors can be evolved or designed for tasks like navigating narrow corridors or avoiding electronic jammers, transferability remains a major hurdle. A behavior optimized for one environment often fails in another.

The challenge lies in creating a decentralized "behavior switcher" capable of running on low-power, single-board computers. Most existing methods require categorical input data or centralized control. Real-world robots, however, operate with "pseudo-continuous" sensor values (e.g., LIDAR distances) and must make local decisions without global oversight. R-HGN addresses this by providing a lightweight, one-shot learning system for real-time environment identification and behavior selection.

Methods

The authors build upon Hierarchical Graph Neurons (HGN), a computationally efficient one-shot learning algorithm. They introduce three key innovations to create R-HGN:

  1. Dynamic Neuron Creation: Standard HGNs pre-allocate memory for every possible input value, which is impractical for continuous data. R-HGN dynamically creates neurons and "gates" only when new sensor values are observed, preventing memory explosion.
  2. Probabilistic Environment Matching: Instead of discrete classification, R-HGN outputs a probability tuple. This allows the system to handle "fuzzy" matches where an environment resembles a mixture of known scenarios.
  3. Temporal and Intra-Swarm Fusion:
    • Temporal Fusion: Probabilities are averaged over $T$ time-steps to mitigate "behavior thrashing" caused by sensor noise.
  4. Intra-Swarm Fusion: Robots broadcast local environment predictions to neighbors. This "belief propagation" allows the swarm to reach a consensus, even if some robots are currently in uninformative regions.
System architecture of the R-HGN framework, showing the training pipeline and the real-time behavior switching loop within the swarm simulator.

Experimental Setup

The system was evaluated on a data-transfer task where a swarm moved 1,000 packets between nodes using three manually designed behaviors: even spacing, wall-circumvention, and jammer-avoidance. The study utilized several simulated environments:

  • Training Set: Six manually designed environments (Env 1.1 to 2.3) with varying wall thicknesses, obstacle layouts, and jammer configurations.
  • Evaluation Set: The original six environments, two "hybrid" unseen environments (Env 3.1 and 3.2), and 50 randomly generated environments to test generalization.
  • Input Features: A 48-component pattern including neighborhood size, sink/source IDs, packet status, and signal strengths (LIDAR and network noise).
Examples of training environments (Env 1.1 and Env 1.2) used to teach the R-HGN system to recognize different spatial configurations.

Key Results and Discussion

R-HGN achieved environment matching accuracy between 92.4% and 99.1% across the designed environments, with $F_1$ scores reaching up to 97.15%. In the 50 random (unseen) environments, R-HGN achieved $\ge 95\%$ of the optimal behavior's fitness in 78% of cases, significantly higher than the 56% achieved by a random behavior selector.

A striking insight is the effectiveness of decentralized consensus. By sharing probabilistic beliefs, the swarm naturally converges on the correct behavior without a central leader. Furthermore, R-HGN allows for spatial heterogeneity; in complex maps, robots near a jammer can switch to avoidance behavior while others continue standard data forwarding.

Conclusion and Future Work

R-HGN provides a robust, decentralized mechanism for swarm adaptation. However, it is currently limited by its predefined behavior repertoire. If a swarm encounters a completely novel environment requiring a new strategy, R-HGN can only select the "least bad" existing option. Future work aims to combine R-HGN with behavior evolution, allowing the swarm to autonomously evolve and add new behaviors to its repertoire upon detecting significantly novel environments, moving toward true life-long learning.

Summary updated

BibTeX
@article{smith2019robotic,
  author = {Phillip Smith and Aldeida Aleti and Vincent C. S. Lee and Robert A. Hunjet and Asad Khan},
  title = {{Robotic Hierarchical Graph Neurons. A novel implementation of HGN for swarm robotic behaviour control}},
  journal = {Expert Systems with Applications},
  year = {2019},
  doi = {10.1016/j.eswa.2021.115675},
}
Figure