Swarmanoid control

Task and environment

The Swarmanoid Task

The task we have chosen is a search and retrieval task. The task can be decomposed into the following sub-tasks that are executable asynchronously and/or sequentially by the robots acting in the environment:

item search
collection/grasping
transport
deposition/organisation

A number of real-world problems can be seen in the terms of some composition of these basic sub-tasks. Therefore, the ability to individually and collectively solve these sub-tasks by the swarmanoid will provide an indirect validation of its potential effectiveness to tackle many real-world problems of interest.

Task Complexity Parameters

The various complexity parameters for the task are shown below.

The swarmanoid will have to search for one or more target objects located either on the 2D floor or in the 3D space of a room. Once the target is detected, it has to be grasped, collected, and transported, either individually or grouped inside a container, to some location that can be randomly selected (e.g., to cluster objects) or selected according to some specific criterion (e.g., to collect the objects inside an assigned shelter). At the deposition location the swarmanoid may be required either to group the objects into unstructured piles, or to organise the objects into some form of structure (e.g., objects could be stacked, or arranged in a pattern).

The default task complexity parameters will involve the swarmanoid transporting the target objects individually to a pre-specified location, and then grouping the objects into unstructured piles.

The Swarmanoid Environment

The swarmanoid will act in an enclosed indoor environment. We make the assumption that the ceiling is ferromagnetic.

Environment Complexity Parameters

Below, we provide a list of environmental parameters that can be altered. A number of different parameter categories have been chosen. For each category the complexity parameters are listed in order of increasing behavioural sophistication required by the swarmanoid. In each category the default parameter is indicated.

Environment Size

Small Room --- Foot-bots can find target in reasonable time frame
Large Room --- Eye-bots are needed to find target (default)

Environment Structure and Obstacles

Simple Structure (One Room), No Obstacles (default)
Obstacles (Walls, Troughs)
Complex Structure (Corridors, Multiple rooms)

Human Interaction

No human present in environment (default)
Human passive --- present in environment. May move, but takes no deliberate action to affect swarmanoid
Human collaborates with swarmanoid
Human disrupts swarmanoid

Target Object Quantity

1 object to be found and retrieved (default)
Many objects to be found and retrieved

Target Object Movability

Single robot can move object
Cooperation required due to nature of object (e.g., too heavy for a single robot) (default)

Target Object Grippability

Gripping easy --- Object designed to be easily grippable (default)
Gripping hard or requires cooperation

Target Object Location

Target on floor
Target raised (e.g., on shelf or table) (default)

Target Object Visibility (size, luminosity, etc.)

Easily detectable --- Object designed to make recognition easy (default)
Not easy to detect

Foot-bots

Vector Field Navigation
	This component is a distributed swarm intelligence control mechanism for distributed foot-bot path formation. The foot-bots form a pattern that globally indicates the direction towards a goal or home location. The foot-bots spread out to form a tree like structure with many branches. Each foot-bot indicates the direction of the adjacent foot-bot in the tree structure that is closer to the root of the tree. Thus, from any point of the tree, a foot-bot can navigate back to the root by following the directional indications of the foot-bots making up the tree. This component could be used for any scenarios in the swarmanoid project where foot-bots are required to navigate between specific locations without the aid of the eye-bots. Research in this component is ongoing and will be described in more detail in a subsequent deliverable.

Adaptive Self-Assembly
	This component is a distributed mechanism to allow a swarm of foot-bots to choose when and if to self-assemble based on the parameters of their encountered task and environment. This component could be used if a group of foot-bots are required to navigate across a-priori unknown rough terrain. If the terrain is simple enough, the foot-bots would be more efficient navigating individually. If, however, the roughness of the terrain requires navigation in assembled formation (e.g., there is a hill too steep for individual foot-bots to navigate) the robots could self-assemble and collectively navigate to their destination. Details of Initial Experimentation

Adaptive Linear Rotation
	This component allows a connected linear formation of foot-bots to adaptively rotate based on the profile of encountered obstacles so as to achieve maximum stability. The mechanism is distributed and relies on LED communication to propagate information along the length of the linear formation. This component could, for example, be used in conjunction with the morphogenesis component if the foot-bots need to navigate over a trough. In this case they could self-assemble into a linear formation (morphogenesis), and then adaptively rotate so as to appropriately bridge the trough. Details of Initial Experimentation

Morphogenesis
	This component is a distributed mechanism to allow groups of robots to self-assemble into particular shapes. Robots already attached to the shape use their coloured LED rings to indicate where new robots should join the pattern. By modifying the rules that govern this process, different patterns can be formed. This component could be used if a group of foot-bots is required to navigate across a-priori unknown rough terrain. Depending on the type of terrain encountered, the robots could choose different shapes adapted to the particular conditions. To cross a trough, for example, the robots could form a line. To cross a hill, on the other hand, a more dense shape might be appropriate. A video on this subject won the best video prize at the AAAI 2007 conference in Vancouver. Details of Initial Experimentation

Negotiated Goal Direction
	This is a mechanism to let the foot-bots negotiate their remembered goal direction to average out individual errors arising from odometry. Each individual foot-bot uses odometry to maintain a direction towards the remembered goal location. However, since such odometric information is integrated, the error made on localization increases with the distance covered. However, if we assume that the errors are normally distributed, by negotiating a common remembered direction it is possible to average out the errors. This mechanism could be used when several foot-bots are trying to transport the target object back to the remembered starting location. Details of Initial Experimentation

Collective Path Selection
	This component extends previous experiments in foraging that involved creation of robots chains so as to create a path between a nest and a prey. Robots could then follow the path and go from nest to prey and vice versa. We extend this framework by considering a case in which there are 2 or more resource sites to be located and exploited by the robots. We expect the robots to exploit the closest resource. If a resource is over-exploited we also expect the robots to balance the load and start exploiting the next closest resource. The requirements could be fulfilled thanks to an implementation of artificial pheromones in chains of robots. Messages are exchanged among robots so as to simulate a flow of ants along the path. When messages are forwarded along the chain, transmitting robots simulate levels of pheromones. The more pheromones there are on a particular path, the more likely it becomes for that path to be selected to route more messages. An amplification takes place and leads to the selection of the shortest path. This component could be used in cases where there are multiple objects clustered into more than one site, and the foot-bots need to efficiently collect the objects from the nearest site. Research in this component is ongoing and will be described in more detail in a subsequent deliverable.

Follow Eye-bot chain
	This component allows foot-bots to follow a chain of eye-bots. It could be implemented in a number of different ways, possibly actively by detecting the location of the eye-bots using a camera, or passively by following a signal sent by the eye-bots. This component could be used in cases where the eye-bots have found the target object and need to direct the foot-bots towards it. Work on this component is in its early stages, as the design and implementation will depend heavily on the finalised eye-bot hardware.

Model Free Fault Detection
	This component uses a back propagation trained neural network to detect robotic faults. The presence of faults is inferred from changes in the flow of sensory inputs and control program behaviour without the need for an explicit model. Research on this component is ongoing. We are currently trying to apply this technique to detect faults in the foot-bots, but the technique should be equally applicable for the hand-bots and eye-bots. This component uses a back propagation trained neural network to detect robotic faults. The presence of faults is inferred from changes in the flow of sensory inputs and control program behaviour without the need for an explicit model. Research on this component is ongoing. We are currently trying to apply this technique to detect faults in the foot-bots, but the technique should be equally applicable for the hand-bots and eye-bots.

Synchronisation
	This component relies on an evolved neural network to create periodic synchronised behaviours whilst avoiding the need for specialised signalling protocols. By simply altering or resetting the phase of their periodic behaviour in response to simple signals the robots are able to efficiently synchronise. This component could be used in a range of circumstances whenever synchronised behaviour is required in the swarmanoid. One example could be that in order to move a heavy object all foot-bots would have to push at the same time, in bursts. Although the initial work on this component has been done on the foot-bots, this component could with little effort be modified to work for hand-bots and eye-bots. Details of Initial Experimentation

Hand-bots

Work on all hand-bot components is still in its early stages, as the design and implementation of all autonomous hand-bot control components will depend heavily on the finalised hand-bot hardware.

Launch Climbing Rope
	This component will allow the hand-bot to launch the climbing rope to the ceiling. The component will need to be able to determine whether or not the launch was successful (i.e. rope correctly attached to ceiling), and be able to take corrective action if the launch was unsuccessful. Work on this component is in its early stages, as the design and implementation will depend heavily on the finalised hand-bot hardware.

Climbing
	This component will be used whenever the foot-bot needs to climb. This autonomous control mechanism will need to take into account the rope that is attached to the ceiling and the surface which is being climbed. Work on this component is in its early stages, as the design and implementation will depend heavily on the finalised hand-bot hardware.

Grasping
	This component will allow the hand-bot to grasp the target object. This component will need to take into account the recognition of the target object as well as manipulation of the hand-bot grippers to successfully grasp the object. Work on this component is in its early stages, as the design and implementation will depend heavily on the finalised hand-bot hardware.

Throw Object Down
	This component will allow the hand-bot to throw an object down to the foot-bots from a raised platform. Work on this component is in its early stages, as the design and implementation will depend heavily on the finalised hand-bot hardware.

Transport Object Down
	This component will allow the hand-bot to transport an object down to the floor from a raised platform. Work on this component is in its early stages, as the design and implementation will depend heavily on the finalised hand-bot hardware.

Eye-bots

Flight Control - Attitude
	Low level stability control of the Eye-bots is vital for the realisation of safe locomotion with the autonomous Micro Air Vehicle (MAV) swarm in constrictive indoor environments. Attitude control, which must maintain a level planar position with no tilt in the pitch or roll dimensions, has been achieved via a high speed PID (Proportional-Integral-Derivative) linear control scheme. Sensory input comes from three high-precision rate gyros that measure rotational velocities in the three relative Euler dimensions and additionally a three-dimensional accelerometer to measure the gravitational component. This has resulted in minimal attitudinal roll and hence minimal disturbances and a stable hover, even under turbulent outdoor conditions. This behaviour will be invisible to programmers of the Eye-bot, in this way preventing potential damage to the platform and allowing programmers to ignore low-level control issues. Research into this component is ongoing and will be described in detail in a subsequent deliverable.

Flight Control --- Lateral Drift
	Due to the fact that the rate gyros mentioned above measure relative rotational velocities and not absolute angular positions, any initial tilt on the platform at take-off will still be apparent during flight resulting in lateral drift, which can have a large velocity. This is because the controller will try to integrate the rotations to this initial reference orientation. In order to prevent this drift, an absolute air-velocity sensor is required to measure lateral displacement. However, to date no such sensor exists that would be appropriate for our hardware. The flying MAV cannot use wheel encoders for odometry sensing and instead must measure air-speed. However, current anemometers have very limited air-speed ranges. Alternatively, relative displacement to stationary objects can be used. One approach being investigated utilises optical-flow inspired by flies to measure relative velocity to the environment. An alternative approach is to make use of the infra-red relative positioning system that the eye-bots will be equipped with. This system will provide a relative position (in the form of a bearing and angle) between neighbouring robots. If one eye-bot is static, via attachment to the ceiling, then the drift of neighbouring flying robots can be controlled by keeping the same relative position. This work is in the preliminary investigative stages. Research into this component is ongoing and will be described in detail in a subsequent deliverable.

Flight Control --- Altitude Control
	An additional critical low level control ability of the eye-bots is autonomous altitude control. That is, the ability to maintain a constant height relative to the ground plane under external disturbances. To this end, the eye-bot has been equipped with an ultrasound distance sensor that perceives the relative distance to the ground via sonar reflections. An additional PID controller is utilised to maintain constant altitude under varying environmental conditions. Results show stable hover with a standard deviation in altitude of only 4.93cm (altitude set-point at 80cm), which is respectable given the limited sensor resolution of 2.54cm (1 inch). Furthermore, by utilising the autonomous altitude control, it is possible to perform autonomous take-off; the eye-bot ascends from the ground quickly towards the set-point with a slight overshoot, before quickly returning to a stable hover at approximately the specified altitude. This altitude control will be available to eye-bot programmers via an interface that will allow a constant altitude (e.g. 1.5m) to be maintained and a secondary function that performs the ceiling attachment. The low-level control algorithm will not be modifiable by programmers of the higher level eye-bot control components thus preventing potential damage to the platform and allowing programmers to ignore low-level control issues. Future work entails the inclusion of a high-precision absolute pressure sensor that measures atmospheric air pressure. This will be used to control altitudes at higher elevations that are out of the range of the distance sensors. This also allows absolute altitude sensing and control regardless of the ground plane. For example, when flying over a desk the distance sensor will sense a sharp decrease in distance, causing the eye-bot to jump up by the same height. However, the air pressure is equal, allowing more accurate altitude control. Sensory data can also be combined via Kalman filtering improving noise resilience. Disturbances in absolute air pressure, e.g. by opening doors or windows, can be controlled by another air pressure sensor placed in a foot-bot on the ground such that the relative pressure difference can be measured. Research into this component is ongoing and will be described in detail in a subsequent deliverable.

Aerial Chain Formation
	This component builds on the work done for the foot-bot chaining component in the SWARMBOTS project. The method of building chains, and chain directionality remains the same. However, the chain is made of eye-bots rather than foot-bots. This component could be used in the initial exploration and search for the targets conducted by the eye-bots.

Direct Foot-bots
	In this component, one or more eye-bots actively direct foot-bots in a particular direction. Development of this autonomous control mehanism will depend on the finalised eye-bot hardware. This component could be used in cases where the eye-bots have found the target object and need to direct the foot-bots towards it. Work on this component is in its early stages, as the design and implementation will depend heavily on the finalised eye-bot hardware.