Fabio Ruini’s blog

Prima o poi diventero’ uno scrittore di abstract. Con quello che mi e’ stato richiesto di scrivere oggi pomeriggio, in piena emergenza, credo di essere arrivato in doppia cifra. E direi che una volta tanto non mi e’ uscito neppure tanto male. Lascio comunque a voi i giudizi.

The work covered by this grant focuses on the use of embodied neural network controllers for MAV (Micro-unmanned Aerial Vehicles) swarms. The goal of the research is to demonstrate how autonomous controllers for groups of flying robots can be successfully developed through computer simulations based on multi-agent systems and evolutionary robotics methodologies.
In the computer simulations we have developed a swarm made of four MAVs has to navigate through an unknown environment, looking for and attempting to neutralize a certain target. In order to make the target inoffensive, a MAV need to perform a low-potential detonation when the distance between itself and the target is lower than a certain threshold. In our hypothesis the MAVs do not have a formalized knowledge about the environment. The only information they can rely on is provided by a satellite system, which informs them about the target’s geographical coordinates. We have tested the swarms’ behaviour in four different experimental setups. First, we have created a simulated environment free of any obstructions. Then we have introduced into the environment many buildings (imitating the buildings’ deployment style typically seen into a Western city, i.e. in our case Canary Wharf, London) that the MAVs are able to perceive during navigation thanks to a set of onboard ultra-sonic sensors. We have then introduced a target able to move, trying to fly away from the MAVs, and a more “robust” one, which requires two contemporary detonations to be neutralized.
The results obtained so far seem to demonstrate the validity of the approach we have chosen to follow. The average percentage of tests concluded successfully for the two easier scenarios is in both cases higher than 85% (respectively 93.46% for obstacle free environments and 87.18% when many buildings are present into the simulated scenario). In case of a target able to move at a speed comparable to the one of a person running, the measured success rate equals to 81.89%. The performance of the neural network-based controllers decreases when the task requires cooperation. In case of a non-movable target that needs two hits in order to be neutralized, the success rate is about 70%. This value becomes minor than 50% when the target is “robust” as well as able to move.
Further investigations on this topic will be focused on the development of forms of communications between the swarms’ members. The idea behind this is that MAVs should greatly benefit from exchanging information when performing tasks requiring cooperation and coordination. At the same time a new 3D simulator is being built in order to develop autonomous controller more easily transferrable on real hardware platforms.

Comments(0)