# Swarms trained in simulation transfer reliably to real robots

*IRIDIA researchers at ULB reported in Science Robotics that constraining a learning algorithm let a 20-robot swarm cross the sim-to-real gap.*

Published: 2026-11-18 · Segment: Learning and Making

![A swarm of small e-puck robots, each with a numbered label and glowing LEDs, spread across a floor arena with a dark corridor and wooden barriers.](https://roboticsdaily.robotshop.com/_astro/iridia-swarm-sim-to-real-science-robotics-2026-11.BiNWJeU7_Z1PVKgx.jpeg)

Image: Mauro Birattari / IRIDIA, Université libre de Bruxelles, via Robohub · press-kit

Researchers at IRIDIA, the artificial intelligence laboratory of the Université libre de Bruxelles, published a study in Science Robotics on how to make robot swarms trained in simulation work on real robots. Robohub reported it on 17 November 2026. The authors include Ilyes Gharbi, Mauro Birattari and colleagues.

Instead of letting the learning algorithm set wheel speeds directly, the team restricted it to choosing among predefined behaviours such as phototaxis, aggregation and exploration. They tested a swarm of 20 e-puck robots on five collective missions, with 250 experiments on physical robots. The constrained approach transferred reliably from simulation and outperformed the leading existing direct motor-control method.

Simpler, more constrained policies transferred better than larger, more flexible ones. The results come from small tabletop robots in a laboratory arena, not from field deployments.

## Sources

- [Robohub (IRIDIA, Université libre de Bruxelles)](https://robohub.org/robot-swarms-trained-by-trial-and-error-in-simulation-keep-working-on-real-robots/)
- [Science Robotics](https://www.science.org/doi/10.1126/scirobotics.adw4624) (primary source)

Read this capsule on Robotics Daily: https://roboticsdaily.robotshop.com/capsules/iridia-swarm-sim-to-real-science-robotics-2026-11/
