Researchers at IRIDIA, the artificial intelligence lab of the Université libre de Bruxelles (ULB), have shown that robot swarms trained by trial and error in simulation keep working reliably when moved onto real robots. Ilyes Gharbi, Mauro Birattari and colleagues report the results in Science Robotics.

Robots trained with reinforcement learning in simulation often fail in the real world, because friction, sensor noise and mechanical imprecision behave differently there. Instead of letting the algorithm set each wheel's speed, the team had it pick among a few tested behaviours, such as moving towards a light, gathering with other robots or exploring, so it only learns when to switch between them.

They tested the method on a swarm of 20 e-puck robots in five group missions, with 250 experiments on the real robots. The swarms kept working after the move from simulation and clearly beat a leading method in which the algorithm drives the motors directly. Simpler, more constrained policies transferred better than larger, more flexible ones.

Birattari, who led the research, said narrowing what the robots can learn keeps the algorithm to choices that remain valid on real hardware. Gharbi said the next step is to test the method on larger swarms, other robots and less controlled settings.

Source: Robohub

Background

IRIDIA

Research lab

The artificial intelligence laboratory of the Université libre de Bruxelles.

Université libre de Bruxelles

University

A French-speaking university in Brussels, Belgium.

Mauro Birattari

Researcher

Researcher at IRIDIA, the artificial intelligence lab of the Université libre de Bruxelles.