Swarm-bots project

This section is devoted to the description of the hole avoidance task. This task has been designed for studying collective navigation strategies in environments presenting holes that need to be avoided. The real s-bot is provided with 4 proximity sensors that are positioned under the chassis, and that can be used for detecting the presence of holes. These sensors are parallel to the tracks, therefore a hole can be perceived only if an s-bot is heading toward it. For this reason, a single s-bot cannot perceive holes positioned on its side and can therefore easily fall into them. In such a situation, the only solution is to rely on a swarm-bot, that can exploit cooperation among the s-bots to perceive and avoid holes. We show how controllers evolved in simulation can transfer to the real s-bots, and we will discuss the results obtained in the real world.

Experimental setup

We evolved hole avoidance behaviors in simulation using the simple s-bot model. In these experiments, the s-bot is provided with the traction sensor, useful for coordinated motion. It also makes use of the ground sensors positioned under the chassis, in order to detect holes. Finally, the s-bot uses its microphones to detect a signal emitted to warn about the presence of a hole (see below). Concerning the actuators, the s-bot can control its wheels and the turret chassis motor, exactly in the same way it does for coordinated motion. Furthermore, a signalling system has been handcrafted in order to emit a continuous tone every time the s-bot perceive a hole by means of its ground sensors. In this way, we implement a kind of reflex action that is triggered when a hazard (the hole) is detected.

In order to evolve coordinated motion and hole avoidance, we have defined three different initial setups: (a) a flat arena with a swarm-bot composed of 4 s-bots connected in line; (b) a flat arena with a swarm-bot composed of 4 s-bots connected in a square formation; (c) an arena presenting holes and open borders, with a swarm-bot composed of 4 s-bots connected in a square formation. The environments "a" and "b" are intended to evolve robust coordinated motion strategies on flat terrain. The environment "c" is devoted to the evolution of hole avoidance. During evolution, the swarm-bot is evaluated in all these different situations.

After a successful evolution, the best genotype obtained from the last generation is selected for testing on the real s-bots. The neural network controller is used on the real s-bots exactly in the same way as in simulation. The values returned by the various sensors are read about every 100 ms, they are scaled in the range [0,1] and finally fed to the neural network. The outputs of the network are used to control the wheels and the turret-chassis motor. The only difference with simulation is a recovery function we added, that is necessary to avoid damage of the s-bots due to excessive efforts by the motors of the wheels. This function constantly monitors the torque applied by the motors of the left and right wheels, and in case the torque exceed a given threshold for a long time, the speed of the wheels is set to 0.

Figure 1: Experimental setup for experiments about hole avoidance. Left: 4 s-bots navigate in an arena presenting "simulated holes", that is, a black band that is perceived as it was a hole. Right: 3 s-bots navigate in an arena with truly open borders. The red lights indicate that the robot is emitting a tone to signal the presence of a hole.

Results

The controller evolved for hole avoidance has been tested on the real s-bots, in order to assess its functionality in the real world. We performed three tests. In the first one, we assessed the capability of a swarm-bot to perform coordinated motion in an environment without holes. We use a swarm-bot composed of 4 s-bots in a square formation. The swarm-bot is positioned in a flat arena, and we measure the distance covered by the group from the initial to the final position. We repeated this test 10 times. In all cases, the s-bots were able to coordinate choosing a common direction of motion, and then moving straight along this direction, while compensating successive mismatches in the direction of the chassis. The obtained performance is very good, the swarm-bot being able to coordinate quickly and to move straight keeping the initial direction in most cases. The average distance covered is 112.2 cm, which corresponds to the 62% of the maximum possible distance a real s-bot can cover in the same amount of time (about 180 cm in average). In few situations, we observed difficulties in the initial coordination phase, which lasted longer than usual. This is due mainly to the fact that sometimes the tracks of some s-bots get stuck and a recovery action must be taken to avoid tracks damage.

In the second test, we placed the swarm-bot in a square arena of 1.5 m side, where the borders are holes simulated using a black band (see Figure 1, left). In fact, ground proximity sensors perceive a black surface in the same way as a hole, because the black colour does not reflect enough infrared light to be perceived by the sensors. This test was performed in order to assess the hole avoidance ability of a swarm-bot before actually testing it into an arena with holes, in order to reduce the probability of physical damages to the s-bots. However, in this experimental phase the dynamics of hole avoidance are different as an s-bot is never suspended over the hole. We performed 20 replication of the experiment, and we observed that the hole avoidance behavior perfectly transferred to real s-bots. In every test, the s-bots where able to detect and avoid the black band.

The third and final test has been performed in a square arena of 1.8 m side, truly presenting open borders. In this case we performed tests using a triangular swarm-bot (see Figure 1, right). Also in this case, we performed 20 replication of the experiments. We observed that in all cases the s-bots were able to move coordinately and to avoid holes. No single fall out of the arena happened, letting us conclude that the hole avoidance behavior efficiently transferred to real s-bots. However, further experiments are required to quantitatively assess the performance drop obtained with real s-bots with respect to simulations.

Hole avoidance

Experimental setup

Results