Ontology-Guided Neural Translation of Human-Swarm Teaming State Space

A primary advantage of reactive uncrewed systems can be attributed to their relatively simple internal decision-making logic requiring little processing power and memory, coupled with their cheap unit cost. When these simple systems are massed together in a swarm, they display collective abilities that far surpass those of the individuals in the swarm. Harnessing the power of swarms is challenging due to their non-deterministic and unpredictable nature. Multiple research fields have developed a suite of techniques to influence or guide swarms to achieve a desired goal, the most promising of which is a bio-inspired control behaviour known as shepherding. Shepherding, derived from observations of biological systems such as dolphins herding fish and sheepdogs herding sheep, utilises a control agent more intelligent and physically capable than the individuals in the swarm. Employing a control agent is imperative for a human to harness the power of swarms to achieve shared goals. Swarm control agents capable of communicating with humans offer teammates increased situational awareness necessary for supervising and intervening with swarm behaviour. This thesis aims to demonstrate the utility of an ontology-guided neural machine translator to foster effective communication between a swarm control agent and a human teammate. This thesis contributes a swarm system-agnostic methodology for effective teaming between humans and swarm control agents. The proposed methodology increases the transparency of the control agent's decision logic through increased contextual relevance and interpretability of its communication. A transparency architecture is presented, and swarming systems are then formalised through ontologies and used to develop a structured language; then, we propose a methodology to fine-tune a transformer model as the communication medium that aims to increase interpretability and explainability. The architecture, associated methods, and transformer model advance the human-swarm teaming research field by successfully interpreting the complex state-space into natural language in the form of English sentences. A combination of novel situation awareness and contemporary natural language processing measures and metrics are used to evaluate our model's translations. The results show that when we apply our methodology, the model's output performance is significantly better than the baseline model. When translating directly from state-space to natural language, our model's output increases the Recall-Oriented Understudy for Gisting Evaluation (ROGUE) score by approximately 60\%, Bilingual Evaluation Understudy (BLEU) score by approximately 60\%, and Bidirectional Encoder Representations from Transformers score (BERTScore) by approximately 30\%. The proposed contributions will enable humans to team with swarms of uncrewed systems with greater scale and autonomy for a more effective human-swarm team.

Grantor dc:publisher

UNSW, Sydney

Year dc:date

2025

Author dc:creator

• Baxter, Daniel ; https://orcid.org/0000-0001-8053-5061

• Human-swarm Teaming
• Human-machine Teaming
• Swarm Control
• Neural Machine Translation
• anzsrc-for: 460202 Autonomous agents and multiagent systems
• anzsrc-for: 460208 Natural language processing
• anzsrc-for: 460205 Intelligent robotics

Statement dc:rights

• open access
• CC BY 4.0
• free_to_read

Licence

• https://purl.org/coar/access_right/c_abf2
• https://creativecommons.org/licenses/by/4.0/

Identifier

https://doi.org/10.26190/unsworks/31443

OAI identifier oai:identifier

oai:unsworks.library.unsw.edu.au:1959.4/105543