Control >> Aggregation >> Probabilistic Control Architecture

Probabilistic Control Architecture

Here, we define a behaviour-based architecture that uses the probabilistic activation of a set of basic behaviours, B. These behaviours, such as the attraction or repulsion from light sources, are designed independently of the pattern to be formed.

In our studies, each basic behaviour creates a mapping from the sensory perception to one of the three actuators (movement, light and gripper) of the s-bot . The basic behaviour set B is partitioned into three, Bm, Bl and Bg, based on the actuator it acts on. These subsets contain basic behaviours that are mutually exclusive, i.e., only one from each subset can be executed at each time step. Behaviours belonging to different subsets can be executed in parallel, since they control different actuators.

The movement behaviours (Bm) are:

• Light Attraction (LA): the s-bot moves in the direction of the light intensity gradient.
• Light Repulsion (LR): the s-bot moves in the opposite direction of the light intensity gradient.
• Robot Attraction (RA): the s-bot is attracted by the presence of other s-bots , detected by the proximity sensors distributed around the s-bot body.
• Robot Repulsion (RR): the s-bot is repelled by the presence of other s-bots.
• Object Avoidance (OA): the s-bot is repelled by the presence of any type of object, walls, obstacle or other s-bots.
• Random Movements (RM): the s-bot moves in a random direction.

The light behaviours (Bl) are:

• Light On (ON): the s-bot turns on the top light.
• Light Off (OF): the s-bot turns off the top light.

The gripper behaviours (Bg) are:

• Gripper Open (GO): the s-bot opens the gripper.
• Gripper Close (GC): the s-bot tries to connect a neighbouring s-bot by closing its gripper.

All three subsets also have a null behaviour that keeps the state of an actuator unchanged. Given the set of basic behaviours, the design of a control system for the swarm-bot to create a certain pattern is then reduced to the design of two parts. First, the context function, h, maps the sensor readings onto a set of contexts that are designed for the particular pattern. This mapping is a categorisation of the sensory perception of the s-bot. Second, the activation probability matrix, P, determines the probability that a given basic behaviour is activated in a given context. In this matrix each row refers to a different context and each column to a particular basic behaviour. The elements of the matrix store the activation probabilities.

Once a behaviour is implemented by specifying the context function and the activation probability matrix, then at each time step the context of the s-bot is computed. Using this context, the probability of activating each behaviour is obtained from the matrix. Then, one behaviour from each subset is chosen based on the activation probabilities at random, and executed in parallel.

The architecture proposed provides a separation of the task-dependent part of the behaviour from the remaining parts that are independent from the task. Therefore, the context function h and the activation probability matrix P fully define the overall behaviour of an s-bot.

Clustering Example

We describe here the probabilistic control architecture applied to the cluster formation. We will describe the context function h, showing for each context a typical situation. Then, we will show the probability matrix that allows to obtain the final behavior.

Context Descriptions

In each figure, the s-bot in lighter color indicates the one the context is referred to.

	Context 1: the s-bot is connected from backwards in this context. Connection sensors are used to determine if another s-bot is connected to its rear or not. Velocity is set to zero, to allow a stable chain. The s-bot turns off the speaker. There is no change in the gripper status, because the s-bot may be at the head of the chain or in the middle.
	Context 2: the s-bot is surrounded by many other s-bots. Camera sensors are used to determine appropriately how close and tight it is surrounded by other s-bots. In this situation, the s-bot must stop its wheel in order to have a more stable cluster.
	Context 3: the s-bot is positioned on the side of the cluster, being surrounded but not in a tight way. In this context, the s-bot has a very high probability for no movement behavior but at the same time, there is a small probability for object avoidance. There are different reasons for this choice. First, using object avoidance the s-bot can find a more stable place in the cluster. Second, the s-bot may rotate around its axis, shifting its context either to Context 2 or to Context 4. In the latter case, if the attraction to sound is not high, the s-bot can escape from the current cluster.
	Context 4: two different situations are considered within this context: a) the s-bot is surrounded by other s-bots as in Context 3, but there are no s-bot in front of it; in this situation the s-bot can move freely.  b) The s-bot does not sense any other close s-bot, in any direction. In this context, s-bot must perform sound attraction behaviour to find a different cluster or to be attracted by the current. Addictionally, random moves and robot attraction behaviours are used to find an "appropriate" place in the cluster, or enable to explore other clusters, escaping from current one.
	Context 5: the s-bot as at least one s-bot exactly in front and very close to it, which may be part of a cluster, or may be a single s-bot searching for a cluster. In both cases, the s-bot tries to escape from that situation trying to find a more stable contact with more s-bots. Thus, obstacle avoidance, sound attraction and random move behaviours are performed.

Activation Probability Matrix :

The activation probability matrix is shown in the following table. The parameter P refers to the probability of switching on the light, which is used to tune the number of cluster and the cluster sizes in the environment. A null value result in no clustering, because there is no attraction to sound. Low values enable the formation of a single, rather dynamic, structure. High value result in the formation of multiple clusters.

Control >> Aggregation >> Probabilistic Control Architecture