Swarm-bots project

In this section we address the problem of how a group of self-assembled s-bots forming a single physical structure can display coordinated movement. In particular, we address the problem of how a swarm-bot constituted by a group of self-assembled s-bots can move as fast and as straight as possible. Given that the orientation of the tracks of each individual s-bot is randomly chosen, s-bots should first negotiate a common direction and then move along such a direction in a coordinated fashion.

Four self-aggregated <i>s-bots</i> forming a linear structure

Figure 1. Left: a single s-bot. Right: four s-bots assembled in a linear structure.

Experimental setup

Experiments have been conducted in simulation by developing a realistic 3D simulator. A swarm-bot consists of a group of four self- assembled s-bots forming a linear structure (see Figure, left). Each s-bot has a rectangular chassis provided with two motorised and two passive wheels and of a cylindrical turret that is connected to the chassis through a "hinge joint" and can rotate freely around the vertical axis with respect to the chassis (see Figure, right). Each s-bot has also a physical link through which it can be attached to another robot along the perimeter of its turret. The link consists of another "hinge joint" that has a rotation axis parallel to the horizontal plane and perpendicular to the line formed by the four robots (i.e. it is rigid with respect to the horizontal plane).

Each s-bot is provided with a sensor placed at the junction between the chassis and the turret that returns the direction (i.e. the angle with respect to the chassis' orientation) and intensity of the force of traction that the turret exerts on the chassis (see Figure). Henceforth this force will be called "traction" for simplicity. A traction force might be due both to the movements of other self-connected s-bots and/or to the movement of the s-bot itself. Notice that the turrets of the s-bots, by physically integrating the different forces that are applied to the robot's own chassis and by the other s-bots, directly provides an indication of the average direction toward which the group is trying to move as a whole. More precisely, it measures the mismatch between the direction where the whole group is trying to move and the direction where the robot's chassis is moving. The intensity of the traction is a measure of the dimension of this mismatch. From the point of view of each s-bot, this type of information is obviously particularly relevant to change the direction of its own track in order to follow the rest of the group or to push the group to move toward a different desired direction. From the motor point of view, each s-bot is provided with two motorised joints controlling the two corresponding motorised wheels.

Figure 2. The sensors encode the angle of the traction force that the turret exerts on the chassis with respect to the frontal direction of the chassis.

Each s-bot's controller is a neural network with 4 sensory neurons that encode the traction. These neurons are directly connected with 2 motor neurons that control the two motorised wheels and the turret-chassis motorised joint. The 4 sensory neurons encode the intensity of the traction from four different preferential orientations with respect to the chassis (front, right, back and left). The activation state of the motor units is used to set the desired speed of the two corresponding wheels and the turret-chassis motor.

The evolving population consists of 100 genotypes each of which encodes the connection weights of a corresponding neural controller. Each genotype encoded the connection weights of a corresponding neural controller that is duplicated 4 times to generate four identical neural controllers for a group of four corresponding s-bots. The fitness function of each swarm-bot, formed by 4 self-assembled s-bots, consists in the Euclidean distance between the centre of mass of the swarm-bot at the beginning and at the end of each trial (each trial consist of 150 cycles). At the beginning of each trial, the chassis of each s-bot is assigned a randomly selected orientation.

Results

By running a set of experiments, we observed that evolved individuals display an ability to coordinate their chassis toward a unique direction that emerges from the negotiation between the individuals, and then move toward that direction by compensating further displacements between individual s-bots Here there is one video of typical evolved behaviours:

We also show how the evolved robots are able to generalise their ability in rather different circumstance by: (a) producing coordinated movements in teams with varying size, topology, and type of links; (b) displaying individual or collective obstacle avoidance behaviours when placed in an environment with obstacles; (c) displaying object pushing/pulling behaviour when connected to or around a given object. See the videos below:

Figure 3. A group of eight assembled robots explores a maze with furrows and holes. The single robots, indicated by arrows get stuck in furrows and holes while the \swb\ succeeds to pass or disentangle from them.

The neural networks evolved to control swarm-bots in a flat arena are robust enough to deal with an environment with furrows and walls. In particular, as shown in Figure 1, while single s-bots get trapped, the swarm-bot succeeds both to pass the furrows and to avoid getting stuck in holes (in the latter case, the collision with the obstacles plays a crucial role, as explained here). Overall, the swarm-bot succeeds to explore the maze by avoiding the obstacles, passing the furrows and disentangling from holes, as the traces left by the robots shown in Figure 1 indicate. The swarm-bot can pass the furrows since the physical links serves as support when some s-bots are suspended over the gap. Once the s-bots at the ``edge'' of the swarm-bot engage in passing the furrows, they are sustained by other s-bots through the physical links. When the s-bots at the edge have passed the furrows, the s-bots at the middle and at the opposite edge can also pass thanks to the support exerted by the other s-bots. This is qualitatively what happens when using the controller evolved only for coordinated motion of a swarm-bot in a flat terrain. The robustness of the controller allows to ignore the small traction forces caused by an uneven terrain, such as the one created by the presence of the furrows. Therefore, coupling the robustness of the controller with the physical support the s-bots mutually exchange, a swarm-bot can efficiently pass over small gaps.

By running other experiments we observed that the ability to coordinate through the traction sensor constitutes an important building block also for developing swarm-bots able to display other and more complex collective behaviours.

Figure 4. Left: A team of 10 s-bots controlled by the neural controllers evolved on team of four individual. The Figure displays how, after some time, the team is able to move almost straight toward a single direction. Right: A team of 12 s-bots forming a circular morphology. The Figure displays how, after some time, the team is able to move almost straight toward a single direction despite neural controllers have been evolved to control only 4 s-bots forming a linear structure

In a first experiment a team of assembled s-bots has been evolved for the ability to find and reach a target light in an large maze environment (see Figure 5). The evolved s-bots in this case are provided with the traction sensor and 4 light sensors. Light and shades are carefully simulated so that s-bots can only detect the target light in absence of obstacles consisting of other s-bots and fixed obstacles such as walls.

Figure 5: A swarm-bot exploring the environment in search of a light target in a maze. The light target is not visible from behind the obstacles. Initially the swarm-bot explore the environment with a random walk (i.e. by following straight trajectories and by changing direction during obstacle avoidance behaviours). However, as soon as one or more s-bots start to perceive the light, the entire swarm-bot move toward the light.

By analysing evolved individuals we observed that, when the swarm-bot is in shadow, it explores the environment by moving in straight lines and by changing direction during obstacle avoidance behaviours. When one or few s-bots start to perceive a light gradient, they start to move toward the light. This behaviour create a traction force that is detected by the other s-bots and that is able to drive the entire swarm-bot toward the light target.

By analysing the behaviour of evolved swarm-bots we observed that they tend to approach the object from "behind" (with respect to the position of the light target) and then start to push the object toward the light target. The traction sensor, in this experiments, play an important role both during the coordination of the s-bots for approaching the object from a convenient relative direction and for the coordination of the s-bots during behavioural phase in which the object is collectively pushed toward the light target.

Figure 6: A swarm-bot pushing an object towards a light target. The object has a light on top, with a colour different from the colour of the light target. At the beginning the swarm-bot navigates towards the objects and reaches a position 'behind the object' with respect to the light target. Then it starts to push the object towards the light target.

These experiments also demonstrate how relatively simple controllers provided with the right type of sensory information and embodied in suitable physical structures might be able to integrate the ability to display several elementary behaviours (e.g. move by coordinating, explore the environment, collectively avoid obstacles, move toward a target object, collectively push the object toward a goal) by selecting the right behaviour at the right time.

Coordinated motion

Experimental setup

Results

References