Control >> Co-ordinated motion

Coordinated motion

In this section, we report the results in which a group of physically assembled s-bots should move as fast and as straight as possible (see Figure 1). The control system that support this basic behavior is based on simple evolved neural controllers and the traction sensor that was first tested in simulation and later included in the real s-bots.

Figure 1: A swarm-bot formed by 6 s-bots connected in a linear formation while performing coordinated motion.

Experimental setup

The controller of the s-bots has been evolved in simulation. A simple model of the s-bot has been developed, and many evolutionary runs have been performed, always obtaining a satisfactory behavior. The controller of each s-bot is a neural network with four sensory units (plus a bias unit) directly connected to two motor units. The sensory units encoded the information obtained from the traction sensor. The activation of the bias unit was set always to one. The activation state of the two motor units is used to set the desired speed of the two wheels motors and the turret-chassis motor.

In each evaluation trial, we tested the behavior produced by the neural controller cloning and downloading it on four s-bots assembled to form a linear swarm-bot. At the beginning of each trial, the orientations of the chassis of the four s-bots were randomly assigned. We observed that the s-bots are able to coordinate by choosing a common direction of motion that emerges from the interactions through the physical connections between s-bots.

Once having obtained a good behavior in simulation, we aimed at the transport to real s-bots. However, before downloading the neural controller into the real s-bots, a problem must be solved. The simulates s-bot and the real one differ in one important detail, that is, in the rotational degree of freedom of the turret. In fact, in the real s-bots the turret has a limited rotation, due to the wiring. On the other hand, in simulation the turret can freely rotate. In order to solve this problem, we engineered a simple solution that allow to use the very same controller and to neglect the hardware limit in the turret rotation. The solution we found is based on a front inversion: the front and the rear of the s-bot's chassis are equivalent, therefore they can be alternatively used, assuming that the sensory information is transformed consistently. In other words, if the control action to a given sensory input corresponds to a forward motion, a front inversion would result to a backward motion.

The rotational constraint problem can be solved by switching each s-bot's front whenever the turret exceeds the 180 degrees limit in either direction, clockwise or anticlockwise. The effect of this change for an s-bot is illustrated in figure 2. The bold arrow indicates the direction of traction. Assuming that the s-bot starts to move by using the first front, the controller first turns the chassis anticlockwise along the direction indicated by the dotted arrow "a" since it perceives a traction from the left. Given that during this turning the chassis reaches the turret-chassis constraint (i.e., the gray wheel reaches the small gray segment) the front is changed. This causes the controller to turn the chassis clockwise along the dotted arrow "b" since it perceives a traction from the right, and as a consequence to move away from the constraint.

Figure 2. Schematic representation of the effect of a front inversion from the point of view of a single s-bot that is turning toward the direction of traction. The circle and the bold open V within the circle indicate the turret and the orientation of the turret. The empty rectangle, the two small gray rectangles, and the two semicircles represent the chassis with its active and passive wheels, respectively. The bold arrow indicates the direction of the traction. The small gray segment indicates the constraint of the turret-chassis degree of freedom (the chassis cannot pass it with the gray passive wheel while turning in either direction). The dotted arrows a and b indicate the direction in which the chassis turns while respectively using the first and second front.

Results

Having solved the turret-chassis constraint problem in simulation, it has been possible to evaluate the performance of the evolved controller on the real s-bots. A number of tests has been performed, in order to assess the quality of the transfer to reality of the strategies developed in simulation. Firstly, 4 s-bots in a linear formation were used, replicating the conditions in which the controller have been evolved. Here, we observed that real s-bots were able to coordinate and to maintain the chosen direction of motion. The performance achieved, measured as the distance covered by the swarm-bot, is about the 85% of the maximum possible (in simulation it was 91%).

Example movies:
• 4 s-bots in line perform coordinated motion.

The neural controller evolved using four s-bots forming a linear swarm-bot is able to generalize to different situations: different number of s-bots and different topology. The following movies gives an idea of the robustness of the evolved strategy to these many situations.

• 6 s-bots in a linear formation.
• 4 s-bots in a square formation.
• 8 s-bots in a star formation.

Figure 3: 4 s-bots connected to an s-toy are coordinately transporting it.

Other interesting generalization features are presented by the evolved coordinated motion behavior. In particular, coordination can still be achieved and maintained when using semi-rigid connections (i.e., connections with a loose grip) that allow some relative movements of the linked robots. Additionally, coordinated motion behavior continue to work when s-bots are not connected directly one to the other, but they are all attached to an s-toy that has to be carried (see Figure 3).

• 4 s-bots connected with semi-rigid links in a linear formation.
• 4 s-bots are connected to an s-toy and transport it.

• 4 s-bots connected in a square foramtion pass over a trough.

Control >> Co-ordinated motion