My research aims to push robot swarms currently operating in carefully controlled laboratory environments out into the real world.

Robot swarms to date are frail systems, unprepared for long-term autonomy. They require hours of learning to adapt their behavior, to recover from faults inevitably sustained during operation. My research aims to remedy this situation by developing an algorithmic framework allowing robots in a swarm to robustly detect faults in each other, and adapt rapidly to unforeseen situations in their environment.

Generic fault detection for large-scale robot swarms


The vertebrate immune system is a complex distributed system capable of learning to tolerate the organism’s tissues even when they change during puberty or metamorphosis, and to mount specific responses to invading pathogens, all without the need of a genetically hardwired characterization of normality.


We have developed a generic fault-detection system for a robot swarm based on mathematical models of the adaptive immune system. The developed system is capable of robustly detecting faults, while adapting itself online to changes in the robot collectives behavior, thus avoiding the need to retrain the system for any new exhibited behavior. It also does not require hard-to-obtain predefined signatures of the various faults robots in a swarm may encounter during operation.

Behavior adaptation for rapid fault recovery

The longdamage_illustration_2-term autonomy of a robot swarm requires robots to rapidly recover from detected faults. In collaboration with colleagues, I developed algorithms that allow robots to recover from faults in no more than two minutes — an order of magnitude faster than state-of-art approaches, and without requiring any self-diagnosis or pre-specified contingency plans. The publication of this result on the front cover of Nature generated a great deal of interest: it attracted over 40,000 views on Nature’s website, was covered by around 40 news outlets (including BBC News, CBC, Washington Post, The Guardian, The New York Times, and BBC Radio), and by over 15 science and technology blogs (including IEEE Spectrum, CNET, Science/AAAS,, Tech Times, and Robotics Trends). The accompanying videos demonstrating robot damage recovery received over 200k views on YouTube.

Evolvability signatures of robot behaviors

Evolutionary robotics is a promising approach to autonomously synthesize machines with abilities that resemble those of animals, but the field suffers fcroppedframe192rom a lack of strong foundations. In particular, evolutionary systems are currently assessed solely by the fitness score their evolved artifacts can achieve for a specific task, whereas such fitness-based comparisons provide limited insights about how the same system would evaluate on different tasks, and its adaptive capabilities to respond to changes in fitness in new situations. To counter these limitations, we formulated the concept of “evolvability signatures”, allowing us to predict the adaptive capabilities provided by a robot’s behavior representation (encoding).

Division of labor in robot swarms

Robot swarms are often required to complete multiple tasks, aAliceArenaClosend are required to divide the tasks among themselves in a distributed self-organized manner. Understanding the mechanisms underlying division of labor in natural swarms such as ants could enable robust robot swarms that efficiently execute multiple tasks.