I am studying how decentralized swarms can move rapidly without losing cohesion when geometry, swarm size, and energy constraints matter. Recent work benchmarks gradient-following in constrained paths and shows where large swarms or sharp turns become limiting; in parallel, I study how local learning can produce energy-saving formation reconfiguration from very limited sensing.

I am studying how communication in robot swarms can move beyond simple signals. By allowing agents to exchange and interpret richer, language-like information, the goal is to support collective environmental perception and social learning in situations where the relevant representation is not fixed in advance.

I use Hebbian learning as a local adaptation mechanism for swarm control. A single learning rule can be shared by all robots while different behaviours emerge online across the swarm, avoiding a separate controller parameterization for every agent and providing a scalable route to learned heterogeneity.

MinionsLLM combines large language models with behaviour trees and formal grammars to translate natural-language task descriptions into structured multi-agent controllers. The aim is to keep language-based control flexible while preserving syntactic validity, explicit behavioural primitives, and a controller representation that can be inspected and adapted.

Fixed-wing UAVs cannot simply wait in place while the rest of a swarm is launched. I studied distributed loitering synchronization: algorithms that align continuously moving aircraft around a loitering circle so that a coordinated mission can start from suitable initial conditions. The work includes a firefly-inspired method that achieves comparable synchronization with much less communication.

With Martin Saska's group, I worked on decentralized UAV swarms that remain coordinated in GNSS-denied, feature-poor environments and under unreliable or absent communication. The work combines onboard mutual perception, relative state estimation, and flocking control, and was validated in demanding real-world experiments, including fast collective flight near the limits of the platforms.

I developed fully onboard swarm control for nano-UAVs to follow dynamic scalar-field gradients in GNSS-denied indoor environments. The robots use local interactions and individual scalar sensing, with obstacle avoidance added without relying on external infrastructure.

I extended automatic design beyond fixed robot bodies by studying controller learning when morphology itself varies. This includes the reality gap in systems where both morphology and control evolve, comparisons of learning methods across previously unseen morphologies, and a model-free method that learns several locomotion skills directly on physical modular robots in only a small number of trials.

I use evolutionary swarm robotics both to engineer collective behaviour and to study how it can arise. In our PLOS Computational Biology work, identical robots evolved division of labour through task partitioning; specialization emerged because environmental structure reduced switching costs, without specialization itself being directly rewarded.

With GESwarm, I used grammatical evolution to synthesize collective behaviours as readable condition-action programs rather than opaque neural controllers. Later work with AutoMoDe-Pomodoro combined evolutionary search with modular behavioural representations, keeping automatic design structured while opening it to a broader range of evolutionary methods.

I developed and analysed mechanisms in which differences in action duration can bias local opinion dynamics toward the better option, even when robots never measure that quality explicitly. This line includes majority rule with differential latency and the k-unanimity rule, tested analytically, in simulation, and on physical robot swarms.

With 100 Kilobots, I studied how a swarm can make fast and accurate binary choices with minimal individual capabilities. The experiments and models show how interaction neighbourhood size controls the speed-accuracy trade-off, providing a way to tune collective decision performance.

I studied how zealots—agents that never change opinion—alter best-of-n decisions, and when informed agents can counteract them. Simulations, mathematical models, and Kilobot experiments show that consensus can still favour the better option provided adverse zealots are not too numerous, and that informed individuals extend this robust regime.

In gradient-driven flocking, only part of the swarm needs access to the environmental signal. Those informed robots bias the motion through ordinary local interactions, allowing the whole group to move toward higher-value regions while most robots remain non-sensory.

In self-organized aggregation, a small proportion of informed robots can bias an otherwise symmetric choice toward a predefined site. The informed robots are not explicit leaders: they follow the same aggregation dynamics as the rest of the swarm, differing only in which sites they are willing to stop on.

Stubborn agents are not always detrimental. In dynamic best-of-n problems, a small persistent minority can preserve alternatives and help the swarm leave an outdated consensus after option qualities change. The same mechanism exposes a trade-off: too many stubborn agents reduce decision performance.

In flocking, a minority of robots with directional information can steer the group while the rest use only local interaction rules. This line started from real-robot studies of self-organized flocking and now extends to collective exploration, where leader-like roles can emerge from local sensing rather than being assigned globally.

I introduced a model of self-propelled agents coupled by spring-like positional interactions but no explicit alignment rule. Below a critical noise level, the group self-organizes into translation or rotation. The mechanism transfers self-propulsion energy toward low-energy elastic modes, providing an alternative route to collective motion based on position interactions alone.

Using the same position-based framework, I studied fluctuations in moving groups and found that velocity and speed correlation lengths grow with the linear size of the group. The resulting correlations are scale-free without requiring the system to be tuned to a critical point, suggesting a direct connection between the interaction structure and large-scale collective dynamics.

My early swarm-robotics work asked what information robots actually need to flock. I showed experimentally that appropriate motion control can produce coherent motion without explicit heading alignment, and studied how directional information and adaptive communication can be introduced while preserving decentralized control.