Research
A collective is only as capable as what each agent can recover from what it sees. My work develops both the coordination across agents and the learning inside each one.
Pillar 1
I develop guidance methods for multi-agent systems, centred on swarm shepherding, where a small number of control agents guide the emergent behaviour of a much larger group. This line of work identified the limits of reactive shepherding (IEEE Access), introduced distributed and consensus-based shepherding, and reframed adversarial patrolling as a shepherding problem (IEEE SMC). To keep the work reproducible, I built and released open-source infrastructure, including a Shepherding Library and a swarm simulation environment that others now reuse and benchmark against.
To drive, the sheepdog positions itself behind the flock centre of mass, on the far side from the goal, so fleeing sheep move toward it. To collect, it positions behind a stray. One sheepdog collects strays first and drives only when none remain.
The circles set the driving and collecting positions. A sheep does not switch behaviour: it always moves on the weighted sum of all its forces, but each force is regulated by distance, so an out-of-range influence, such as a sheepdog beyond the threat radius, contributes zero and drops out of the sum.
Pillar 2
Real-world data is partial, noisy, and high-dimensional. I design learning architectures for exactly this, in a sustained line of work on gated autoencoders: from Gate-Layer Autoencoders for incomplete EEG signal recovery (IJCNN), to the Weighted Gate Layer Autoencoder, a learnable gating mechanism that models inter-variable dependencies and reconstructs missing signals (IEEE Transactions on Cybernetics), to gate-control mechanisms that make this selection more precise (Sensors).
A gate layer chooses which variables carry the reconstruction, recovering what is missing.
Where it meets
I am extending these foundations toward collaborative learning across agents that differ, and on-board perception under tight compute budgets, so that distributed perception and decision-making hold up on resource-constrained platforms.
Applied autonomy
These methods are tested and translated through funded projects with industry and research partners, moving results from concept toward real-world readiness.
Selected results
A learnable gate layer switches input variables on and off during training, so the model learns which variables depend on which. When channels drop out, reconstruction error stays low even with up to ten channels missing at once, where a standard autoencoder degrades sharply.
IEEE Transactions on Cybernetics, 2021. doi:10.1109/TCYB.2021.3049583
A systematic study of what makes swarm guidance hard, run with 100 sheep and 1 to 10 control agents in a 50x50 field. Using the paper's obstacle-density thresholds, it maps a reliable-herding region and a collapse region, with the transition shifting from roughly 0.2% density for one control agent to about 5% for ten.
IEEE Access, 2020. doi:10.1109/ACCESS.2020.3037325
Adversarial patrolling recast in the continuous domain: attackers are a swarm drawn to a protected area, and a single defender repels them using force-based behaviours. A one-step look-ahead picks among driving, collecting, intercepting, and patrolling to hold the swarm at a safe distance.
IEEE SMC, 2024. doi:10.1109/SMC54092.2024.10832074
Air traffic control recast as shepherding: the controller moves a single logical shepherd in the digital airspace, and its force rotates each aircraft’s intended flight plan rather than steering positions directly. The influence propagates from aircraft to aircraft, so one controller redirects a whole fleet without per-aircraft commands. Forces sheep need but aircraft don’t are absorbed by the autopilot, flight plan, and onboard collision detection. Demonstrated through a new ATC interface in the ATOMS air-traffic simulator.
Springer, Shepherding UxVs for Human-Swarm Teaming, 2021. doi:10.1007/978-3-030-60898-9_11
For EEG, the order in which channels are gated matters. Embedding the electrode layout into the gating scheme, by anatomical order, by lobe, or by hemispheric symmetry, improves reconstruction over random gating and keeps accuracy high even when many channels are blocked.
Sensors, 2025. doi:10.3390/s25113389
Two translators produce different translations of one source, so the data is paired and the informative signal is the change between the paired feature values. We framed this as a Pairwise Comparative Classification Problem and modified the C4.5 decision tree into PWC4.5, which splits on a relationship over a pair's values rather than treating instances alone. On translator identification, standard C4.5 reached 52.12% while PWC4.5 reached 80.23%.
ACM TALLIP, 2016. doi:10.1145/2898997
PLOS ONE, 2019. doi:10.1371/journal.pone.0211809
Funded research
I am a lead and co-investigator on competitive grants totalling over AUD 1.2M, spanning applied autonomy with industry and research partners, an industry-supported HDR scholarship, and shared research infrastructure for swarm and distributed-intelligence experimentation.
The group
Ten higher degree research candidates and completions, working across both pillars. Meet the group →