• No se han encontrado resultados

FORTAZ IN D5W SOLN 2-5 GM/50ML-%

CARDENE IV SOLN 40-5 MG/200ML-% [nicardipine

FORTAZ IN D5W SOLN 2-5 GM/50ML-%

The subject of swarm robotics follows closely on from swarm intelligence. It has been defined by S¸ahin as:

The study of how large number of relatively simple physically embodied agents can be designed such that a desired collective behaviour emerges from the local interactions among agents and between the agents and the environment [78]

“Relatively simple” means that an individual robot is incapable of performing a partic- ular task, but the same task is achievable through the collaboration of multiple robots. It can be seen from this definition that swarm robotics and swarm intelligence are very closely related subjects; they both concern the mass collaboration of simple agents to achieve some global goal and the investigation of how such emergent behaviour is pos- sible. In actuality, a swarm robotic system is a form of artificial swarm intelligence, as such a swarm robotic system displays the characteristics of a swarm intelligent sys- tem: decentralised control, self-organisation, asynchrony, local interactions, stability and adaptability.

In swarm intelligence research the biological system is modelled and simulated to get a better understanding of the mechanisms behind the swarming behaviour. Compara- tively, swarm robotics can be seen as the modelling and simulation of swarm intelligence theories using embodied agents, by designing the agents in the system on a microscopic level to try to produce macroscopic emergent behaviours. However, the goal of swarm robotics is not always to accurately copy the inspiration but to achieve the emergent properties of the inspiration for our own benefit. For example, it would be worthwhile for us to be able to make robots that can bridge a gap or collectively push heavy items from one place to another as the Weaver ant does [38, 49]. In both these examples the biological system is used as inspiration for the programming and behaviour of the robots. The methodologies evolved by the ants are heavily abstracted to fit the real- world application which the swarm robots will be used in. The swarm robots may still be seen to be obeying the same rules as the ants and as such are modelling the ants. However, unlike conventional modelling as described in section 2.4.2 what is important is the swarm’s ability to perform the task. As long as the task is adequately performed, how closely the robots match the biological system is irrelevant.

An engineer with a problem to solve does not have to be concerned with the biological plausibility: efficiency, flexibility, robustness and cost are possible criteria that an engineer could use. [14]

In our work, swarm robotics has two purposes: The first is to learn more about emergent swarm behaviours through attempts to replicate them, the second purpose is to engineer a multi-robot system that achieves some beneficial task through swarming behaviour.

3.3.1

Benefits of Swarm Robotics

Swarm Robotics as a Modelling Tool

The advantages of modelling have already been discussed in section 2.4.1. As a sim- ulation tool, swarm robots have the advantage that they are an embodied simulation platform. The advantage of using an embodied simulation over a computerised one is that on a computer it is hard to accurately emulate real-world physics. Simulating something like a ball being knocked across a testing arena would be effortless with an embodied system, but in a computerised simulation factors such as the ball’s physical properties, the friction and gradient of the surface and the strength and direction of the push need to either be measured or approximated. Furthermore the resulting path of the computerised ball would not be completely accurate or reliable. We can never fully know all the factors which affect the ball, much like we can never fully know all the factors affecting the agents in a natural swarm (this is properly discussed in section 2.4), but at least with an embodied simulation we do not need to concern ourselves with modelling the physics of the environment. By placing our simulation within the real rather than approximated world, the physics is modelled for us.

Another advantage of embodied simulation is that the real-world is far richer, in terms of information content, than a computer simulation. In a physically homogeneous set of robots there will be different imperfections within each individual robot. For example actuators may not respond in exactly the same way to a command on one robot when compared to another robot. Similarly there may be noise in the sensor readings, a proximity sensor could easily give two different range measurements for the same distance. In computer simulation these differences can be approximated by adding noise, on the sensors and on the motors. An infra-red proximity, for example,

has a minimum distance, any objects that are closer than this distance will result in wildly inaccurate readings. Objects within the minimum and maximum ranges will give more accurate, but still noisy, readings. It is not enough to model these sensors with a constant noise level, and we cannot assume that items will not get closer to the sensor than its minimum range. Using embodied robots we already have noisy sensors and robots, with a computer simulation there needs to be careful consideration and testing to create a realistic and reliable model.

In addition to the richness of the robot there is richness in the environment which we place our robots in. We have already discussed this richness in terms of the world’s physics, but there is also a wealth of information in the real world which the robot needs to be able to filter out. If a robot with a camera is tested in a messy room, there is much more information presented to the camera than if the room was empty and painted completely white. The robot must work harder to extract only the relevant information from its camera; this can be both good and bad. The benefit is that the robot is much more robust to sensor noise and consequently can better function in a wider range of environments; the downside is that this extra functionality is considerably more difficult to create.

Brooks summarises the advantages of embodiment over simulation in the phrase: The world is its own best model. [15].

Swarm Robotics as an Application

Beyond the realm of modelling, swarm robotics is an attempt to create robotic systems that have the advantages of natural swarm intelligent systems (as outlined in section 3.2.1), and are capable of achieving some collective task in a robust, flexible and scalable way [78].

Robust Robust swarm behaviour is a result of decentralised control: the robot swarm has no global controller to be a single point of failure for the entire swarm. Additionally the robot members of the swarm are simple, both mechanically and functionally, so there is less potential for failure. The robot agents are also performing similar behaviours with high redundancy and so the loss of one swarm member should not affect the functionality of the swarm as a whole. Trianni [89] makes the point that decentralisation and redundancy must both be present for a robust collaborative behaviour. Trianni uses the example of a factory production chain: there is decentralised control in that each robot does the same simple task repeatedly, but if one robot were to fail the entire production chain would fail because their tasks are all different and rely on one another to be completed. There is no redundancy. Winfield and Nembrini [95] give three more desirable features of a robust robotic swarm:

• It should be tolerant to noise and uncertainties in the operational environment. • It should be tolerant to the failure of one or more robots without compromising

the desired overall swarm behaviours.

• It should be tolerant to individual robots who fail in such a way as to thwart the overall desired swarm behaviour.

These criteria link back to the properties of stability and adaptability that are displayed by natural swarm intelligence systems. A truly robust swarm of robots should be

able to maintain its functionality despite failings within the swarm or changes in the environment outside the swarm. However, Winfield and Nembrini [95] have shown that not all robotic swarms are robust to failures. In their paper they show a swarm performing an aggregation and taxis task, where all the members of the swarm come together and move away from a beacon. Aggregation is achieved by each robot using short distance wireless link to broadcast its presence, and trying to maintain the number of robots within wireless range above some threshold β. To achieve taxis the robots have a sensor which detects when a beacon is shining upon them. When a beacon is detected the robot increases β to ∞ so that there is differential movement in the swarm, leading to taxis behaviour. Then Winfield and Nembrini partially failed a small number of robots in the swarm, testing the effects of failing the motors, wireless communication, beacon sensor, proximity sensors, control system and finally the entire robot. They found that the robot swarm was robust to failure of the beacon sensor and to the entire robot failing. Losing the wireless connectivity of a robot caused it to become lost but the swarm was also able to withstand the change. Losing the proximity sensors caused collisions but was also withstood. A control system failure was simulated by the motors becoming stuck either moving forwards or turning on the spot. The former caused the robot to become lost, an effect the swarm was able to adapt to. The robot becoming stuck turning on the spot caused the same effect as failing the motors. The robot would become fixed in one spot but it would continue to wirelessly communicate with other robots and would consequently “anchor” the whole to the spot where the robot was stuck, thereby causing the entire swarm to fail in its task.

What this study ultimately shows is that a robot swarm can be robust to some major failures; the loss of an entire robot agent is easily adapted to, and this is a considerable advantage of swarm robotics. However, for a truly robust swarm it is important to design the swarm behaviour so that it can withstand partial failures and to thoroughly test the swarm’s reliability under such circumstances.

Flexible S¸ahin [78] describes flexibility as another advantage of swarm robotics. He defines this as being a swarm’s ability to coordinate its behaviour to perform different tasks [79]. Just as an ant colony is able to distribute agents between the tasks of foraging, cooperative transport, fighting off attackers and so on, so too must a swarm robotic system be able to perform different tasks as is required by its environment. Trianni [89] also describes flexibility as being a characteristic of a swarm robotic system, however, unlike S¸ahin, Trianni defines flexibility as being adaptive to environmental changes instead of a diversity of tasks performed. Both definitions are similar though, in that they both require an appropriate response to the environment from the robot swarm. The amount of different behaviours required from the swarm is perhaps therefore a reflection of the complicatedness of the environment, a more difficult situation requiring a larger repertoire of responses from the swarm.

Scalable The final major benefit of swarm robotic systems is that they are scalable. This means that it is possible to add or remove robots from the swarm and it will still continue function. Natural swarm systems can coordinate millions of different agents, so it should also be possible for artificial swarm systems to coordinate millions of robots. This is a beneficial characteristic because a greater number of robot agents means that the swarm has more resources available to complete its task and it has a higher redundancy and so is better able to cope with the loss of agents.

and Nembrini [95], shows that scalability is not always a guaranteed benefit of swarm robotics. Winfield and Nembrini [95] show that a swarm can be fragile to partial faults. The work by Winfield et. al. not only confirmed this result but also found that increasing the number of robots in the swarm does not fix the problem but, in fact, makes it less robust: a larger swarm was observed to take longer to recover from a partial fault, and an increased number of robots were lost during the recovery. Consequently, scalability may not be as straightforward as adding more robots to the swarm. Careful consideration needs to be given to the process of increasing the swarm size if the aim is to create truly scalable systems.