Mechatronics 13 (2003) 195–228
Towards robot autonomy in the natural world: a robot in predator’s clothing John Greenman a, Owen Holland b, Ian Kelly b, Kevin Kendall c, David McFarland b, Chris Melhuish b,* a
Faculty of Applied Science, University of the West of England, Bristol BS16 1QY, UK b Faculty of Engineering, Intelligent Autonomous Systems Engineering Laboratory, University of the West of England, Bristol BS16 1QY, UK c Department of Chemistry, Birmingham University, B15 2TT, UK Received 26 June 2000; accepted 22 March 2001
Abstract This paper reports the progress on an extensive inter-disciplinary exercise, namely building a self-sufficient robot, that can catch slugs and obtain energy from them. This project involves engineering design and manufacture of the robot, including its propulsion, slug-catching and navigation mechanisms; the development of behavioural strategies for optimal deployment of time and energy; the development of a digester, capable of breaking down slugs into a usable chemical medium; and the development of fuel cells for extracting electrical energy from this medium. Ó 2003 Elsevier Science Ltd. All rights reserved.
1. Introduction Robots require energy, which is usually supplied by their human owners. In the future we hope to see self-sufficient robots carrying out useful tasks without human supervision. Such robots would be autonomous in the sense that they obtain their own energy, and in the sense that they control their own behaviour. The concept of autonomy has arisen in various contexts, including artificial life [13,26,27], biological robotics [29,31], and practical engineering, which is our concern here. The primary aim of the ‘SlugBot’ project is not one of pest control, but to investigate the
*
Corresponding author. Tel.: +0117-344-2539. E-mail address:
[email protected] (C. Melhuish).
0957-4158/03/$ - see front matter Ó 2003 Elsevier Science Ltd. All rights reserved. PII: S 0 9 5 7 - 4 1 5 8 ( 0 1 ) 0 0 0 4 5 - 9
196
J. Greenman et al. / Mechatronics 13 (2003) 195–228
engineering of a self-sufficient robot engaged in the real world pursuing a real world prey. Slugs are a pest and therefore a suitable prey for this study. Energy autonomy concerns the ability of the robot to obtain its own energy. The first person to demonstrate that this could be done in laboratory robots was Grey Walter [47,48], whose electromechanical robot used phototaxis to home in on a recharging station when its on-board energy supply reached a low level. In more recent years, this limited form of energy autonomy has been demonstrated many times. In addition, partial energy autonomy, in the form of solar-powered vehicles has been employed in lawn mowers, space vehicles, etc. Solar power has the advantage that the energy is readily transformed into an electrical supply of energy to the robot. Other sources of energy, such as plant and animal organic matter, pose the problem of converting the source energy into a form that can be used by the robot. This paper addresses this problem by attempting to design and construct a robot system capable of sustaining itself for periods of time by foraging for a natural energy source, its food, and converting this into a usable energy form to power the robot. Such an endeavour forces us to confront real issues which will demand technically challenging engineering solutions [14,18–20]. Let us briefly consider some of the main issues: a robot needs to be mobile in order to forage for food. This food supply may only be available at certain limited times and locations depending upon climate, time of day, seasons and so forth. Much depends upon the nature of the food. Static vegetation necessitates a different foraging strategy than that required for the hunting of mobile prey. By whatever means the food is found it is then to be converted and stored into a usable form; the conversion process constituting some kind of artificial metabolism. One could envisage a system incorporating the elements of a mobile robot and an energy conversion unit. They could be combined in a single robot or kept separate so that the robot brings its food back to the ‘digester’. One option is to employ on-board artificial metabolism on a single robot [50]; another is to use multiple robots each with its own digester. Yet another system could employ an external central metabolism station (central digester) supplied with digestible material by one or more robots. Whatever combination is employed, it will need to reflect the energy budgets, metabolic efficiencies, type of food, physical size of digester and so on. In this study, we have chosen to physically separate the robot and digester system. A fermentation unit is likely to be too large and heavy to be carried around by a robot. A strategy similar to that of leaf-cutter ants is therefore employed; ants bring back leaf cuttings to a ‘garden’ where the organic material is converted into a fungus which the ants eat. In this way the robots forage for slug material and bring it to a central ‘digester’ which converts the biomass into electricity which powers the robots. It is envisaged that a central digester system will comprise of a two stage anaerobic fermenter and fuel cell system. We chose slugs as the robot prey, for the following reasons: the prey should be sufficiently plentiful and be of a sufficient energy density to justify the energy expenditure involved in searching, catching, transporting and converting to biogas. The prey should not be capable of rapid movement since this would require the robot to expend considerable energy in catching the prey. An in-
J. Greenman et al. / Mechatronics 13 (2003) 195–228
197
vertebrate pest already subject to lethal control measures would conform to ethical considerations. Slugs are slow moving and abundant on UK agricultural land; especially Deroceras reticulatum [38]. They represent a considerable threat to vegetation, perhaps doing most damage to crops requiring a well-cultivated seedbed, such as potatoes or wheat [10]. Their destructive power is countered by UK farmers who spend over £20m per year on chemical control measures [11]. Our first trials will employ a single robot that will hunt for slugs and bring them back to the digester. The digester will provide power for the robot to carry out its functions. From initial calculations it is unlikely that a single robot could supply sufficient biomass to the converter to generate sufficient energy to cover the requirements of the robot and the digester itself. One possibility, in the longer term, is to have a solar-powered digester or a multi-robot system employing a large central digester. Taking all the above considerations into account, we have designed a wheeled, battery-powered mobile robot that is autonomous in the sense that it can locate and home in on its charging station, dock and recharge its batteries. This, in itself, is not the first [34,47], but our robot is intended to be energy self-sufficient, because instead of the electrical energy being provided by humans, it is to be provided by the robot itself. The robot is equipped with a long arm (see Fig. 1), with a grabber at the end. In the ‘palm’ of the grabber is a camera by means of which the robot can detect the presence of slugs at night. The robot can pick up slugs and deposit them in its
Fig. 1. The robot.
198
J. Greenman et al. / Mechatronics 13 (2003) 195–228
‘pocket’. When the robot needs to recharge its on-board battery, it homes in on the recharging station and docks. While it is recharging its battery, it unloads the slugs which it has collected. These enter a ‘digestive system’ which is part-and-parcel of the charging station. The product of digestion is a biogas that is fed into a fuel cell, which produces the electrical energy that ultimately powers the robot. The remainder of the paper is divided into four main sections. Section 2 addresses the issues associated with energy autonomy and its impact on the requirements of a self-sufficient robot. These ideas are explored further in Section 3 which discusses the nature of generating fuel from biomass. Drawing upon the ideas set out in the previous two sections, Section 4 focuses upon the construction and design of a slugcatching robot. Finally, Section 5 sets out and stresses the major points of the study.
2. Energy autonomy Like animals, robots may have differing degrees of energy autonomy. The different types of energy autonomy found in animals have their parallels in robotics [28]. We here concerned with the energy obtained from the natural environment, and we can make a start by looking at foraging in animals. 2.1. Foraging in animals Natural selection favours efficient foragers, and most animals are extremely adept at searching for, and harvesting, food. Different species employ different foraging methods, some searching for food, some lying in wait for prey, some grazing, etc. Thus some species have a high rate of energy expenditure while foraging, but spend little time foraging, while other species have a low rate of energy expenditure, but spend a lot of time foraging. The rate at which energy is obtained depends upon the availability and accessibility of the food. These determine the rate of return on foraging. Efficient foraging is usually a matter of trade-off between competing priorities. These may include energy gained vs energy spent, energy gained vs risk of predation, energy gained vs losses to rivals, etc. The trade-offs involved in foraging have been extensively explored, both theoretically [44] and experimentally [23], see [28] for a review. In order to obtain food, animals have to expend energy. They also have to expend other valuable commodities, such as time, water, heat, etc. Given freedom of choice, animals should trade off among the alternative possible behaviours and perform the one that it most beneficial in terms of fitness. However, the animal does not always have freedom of choice because of constraints operating on and within, the system. Such constraints upon animal foraging may include the energy available to spend on foraging, time available, limitations on efficiency due to handling prey, etc. Such constraints have been the subject of a considerable amount of investigation (for reviews see [23,32,44]). Animals usually consume their food on the spot, but if an animal has a nest, or comes from a colony, it may take food home. Such animals usually make an outward
J. Greenman et al. / Mechatronics 13 (2003) 195–228
199
journey, spend some time searching and then make a return journey. This type of foraging is called central place foraging, and is subject to optimality considerations that are different to those of ordinary foraging [35], for reviews see [23,28]). The large body of knowledge about animal foraging provides a useful starting place for the design of behavioural strategies for foraging robots, and our approach is predicated on this basis. 2.2. Self-sufficiency in robots To be truly self-sufficient, an agent must exhibit both behavioural stability and market viability. Behavioural stability implies that the agent does not succumb to irrecoverable debt of any vital resource. The vital resources are those that enable the agent to carry out its design tasks, and may include energy, time, tools, etc. Irrecoverable debt is not simply a question of running out of a vital resource, but may include debts, the repayment of which engenders other debts. Behavioural instability may occur if the agent runs out of a vital resource, or if the servicing of debts takes so much time and/or energy that the agent is not able to carry out its design tasks. Market viability amounts to pleasing the robot’s employer [28]. The employer will be satisfied if the agent is behaviourally stable, provided that the agent is also able to perform the tasks that it is designed to perform, and provided also that the running costs are acceptable. Self-sufficient agents must have a degree of autonomy, because they must have the freedom to decide for themselves when to refuel, when to perform certain activities, etc. (see [29,31]). The degree of autonomy usually equates with the number of basic resources that the agent has to manage. Simple self-sufficient agents are selfish in the sense that they manage their own resources regardless of other agents that may be operating in the ecosystem. Such agents usually have to trade off between refuelling activities and activities designed to please the employer (call this working). There are two basic resources that must be provided by the robot environment, if the robot is to be self-sufficient and economically viable. These are energy E, which the robot must be able to obtain in some way, and M which can be obtained by working. M can stand for memory of the amount of work done, merit points for work done, or money, or market viability. Each time the robot does a certain amount of work (i.e., fulfils part of the task that is useful to the owner) it earns a unit of M. When we consider a single self-sufficient robot, it is evident that it should perform a basic cycle of activities to maintain viability. The robot goes through a cycle of work – find fuel – refuels. When working, the robot gains M and E is lost. At some point the robot breaks off work and goes to find fuel. This also leads to a reduction in E, but what about M? The answer to this question depends upon the attitude of the owner of the robot. M represents the utility of the robot’s work from the point of view of the owner of the robot. There are three basic possibilities, as outlined in Fig. 3. If the owner is primarily interested in robots that spend as much time as possible doing useful work, irrespective of energy expenditure, then the owner will not gain
200
J. Greenman et al. / Mechatronics 13 (2003) 195–228
Fig. 2. Basic cycles AB.
utility from the time that is spent not working. This means that the robot should be designed so that M declines during unproductive time, as shown in Fig. 2(a). If the owner is concerned about energy expenditure on activities that are not productive, then M should decline during that period of the basic cycle, as shown in Fig. 2(b). If, on the other hand, the owner is concerned to minimise the energy expenditure across the board, then it makes sense for the robot to pay for its fuel. In Fig. 2(c) M is earned during work and spent during refuel. In other words, M is like money which the robot earns by working and spending off fuel. The stability of the basic cycle depends upon two main types of factors; the nature of the environment, and the decisions made by the robot. Thus the system can become unstable because the environment is just too difficult for the robot to cope with. An animal equivalent might be an environment in which food was very scarce. The system could also become unstable simply because the robot made bad decisions, such as ignoring an opportunity to recharge (i.e., when near the recharging station). Thus behavioural stability and decision-making are closely connected. Examples of stable and unstable basic cycles, obtained from experiments, can be found in [34]. 2.3. SlugBot capability The aim is to design a robotic system in which a solar-powered sessile station is established in the field to serve a number of mobile robots. The station should be capable of the following:
J. Greenman et al. / Mechatronics 13 (2003) 195–228
201
(1) acting as a recharging station for a number of foraging robots, although this paper will be concerned with the situation for only one robot, (2) receiving the robot prey and converting it into electrical energy, (3) carrying out tasks on its own accord, such as surveillance, etc. The robot ‘predator’ should be capable of the following: (a) moving around free from outside control, (b) finding and catching slugs, its prey, (c) visiting the station to deposit its prey and recharge its batteries. 2.4. The behaviour repertoire The behaviour repertoire of the station will consist of a number of entirely compatible activities, such as recharging robots, turning the solar panel, carrying out other tasks, etc. Therefore the repertoire of the station is of behavioural importance only insofar as it causes variations in energy gain and enables the robot to unload slugs. The behaviour repertoire of the robot will consist of the following incompatible (in terms of time) activities: Recharging. The robot recharges its batteries at the station which will allow many robots to recharge at the same time – which may not be physically next to the unloading dock. Unloading. The robot unloads its prey at the station. We envisage that only one robot will unload at any one time. Unloading will take minutes whereas recharging will take many hours. Docking. The robot docks at the station for the purpose of unloading and recharging. Searching. The robot moves around searching for a suitable location for catching slugs. Scanning. The robot scans for slugs, while stationary. Catching. The robot catches a slug, and puts it in its pocket. Homing. The robot goes home by the most direct route. Resting. The robot becomes immobile, not doing any of the above. 2.5. The SlugBot basic cycle The most simple basic cycle for the slug-catching robot is made up of the following cycle of activities: recharging, searching, scanning, catching, homing, and docking. We now have to consider which type of basic cycle is appropriate. Of the three types illustrated in Fig. 2, (a) is not relevant here, because energy expenditure by the SlugBot is important, and if the SlugBot cannot meet all it own energy requirements, extra energy will have to be supplied, and this will be costly and inconvenient. In the case of possibility (b), where the owner is concerned only about the energy expenditure on activities that are not productive, we have to remember that the
202
J. Greenman et al. / Mechatronics 13 (2003) 195–228
robot can rest during unproductive times. During resting its energy expenditure will be very much reduced. The owner of the SlugBot should want to minimise energy expenditure across the board, and therefore basic cycle type (c) is the most appropriate. As we can see from Fig. 2(c), during recharging the robot has a net gain of energy E. During all other activities it has a net loss of energy. During recharging the robot has a net loss of M, while during searching, scanning and catching the robot is considered to be working, and has a net gain in M. During homing and docking the robot neither gains nor loses M. Once the robot is at the recharging station, it can take energy on board, provided it ‘pays’ for it in units of M. In other words, M is like money. There are two questions of policy that now have to be considered. (i) How much M should the robot pay per unit of energy gained? (ii) Should the robot be allowed to run up a debt at the charging station. To some extent the answers to these questions depend upon the situation that the charging station is in, and the answers to these questions are a matter of station charging policy which is yet to be decided. Note that the robot does not have to recharge fully. If it senses that there is a good slug harvesting opportunity, then it could decide to break off recharging and start foraging. Once the SlugBot starts foraging, its success depends largely upon the availability of slugs at the site where it is foraging. High availability will lead to a high income (in terms of M) for the robot, provided the robot forages efficiently. Obviously, the robot should stop foraging if, on average, the energy expended is greater than the energy gained. This might happen if slug availability is very low, or its foraging effort (i.e., energy expended per unit time) is very high, due to bad weather or bad terrain. If the robot does decide to stop foraging, then it must take one of the two options: (i) go home, or (ii) rest. The former will involve high energy expenditure (depending upon the distance to the station), while the latter involves low energy expenditure. Provided the robot remains stationary, while resting, and just expends enough energy to keep its sensors going, then it will expend very little energy. The disadvantage of resting is that the cost of going home still has to be paid in the future, because eventually the on-board energy will run out. 2.6. Functional analysis Having outlined the basic cycle for the SlugBot, and it is clear that, from the functional point of view, there are a number of trade-offs involved. Clearly the robot must not spend too much time and energy on work, or it may be unable to find the station before it runs out of fuel. The robot must decide when to stop working and start looking for the station, if it does this too early it will not be working to maximal efficiency, and if it leaves it too late it may run out of fuel. Similar arguments apply to the decisions to stop recharging and start work, or stop foraging and take a rest. We can now imagine an optimal cycle, in which the robot changes from one activity to another at the optimal point. Obviously, the degree to which the robot is able to make optimal decisions will affect its behavioural stability [34].
J. Greenman et al. / Mechatronics 13 (2003) 195–228
203
From an optimisation viewpoint (see [32]) the optimality criteria are functions of the state variables, in this case E and M. Taking E first, it is clear that there is a lethal boundary on this state variable, because if the robot runs out of fuel it is considered to be dead. In such cases the optimality criteria take a quadratic form [32]. In other words, the risk of running out of fuel increases as the square of the energy deficit. Turning now to M, it will be apparent that if the robot has no M it can buy no fuel. This is akin to having no money with which to buy food. So, in the case of the type of basic cycle that we have chosen (Fig. 3(c)), there is also a lethal limit on M, and we can expect the relevant optimality criterion to be quadratic. Had we chosen a different type of basic cycle, we could expect the robot to get the sack should M ¼ 0, since this is an indication that the robot has done no useful work. So, as a preliminary working assumption we can expect the optimality criteria to take a quadratic form. What we are doing here is analysing the behaviour of the robot in terms of its basic function. So far we are saying that it is a basic function of the robot to do useful work without running out of energy. If we want the robot to be as energyefficient as possible, then we are also saying that a basic function of the robot it to trade off the risk of death from E ¼ 0 against that from M ¼ 0. There are also other trade-offs to consider. While foraging, there is a forage-rest trade-off, and while recharging, there is a recharge-forage trade-off. Basically, what the robot should do is to take the option which has maximum utility at the time. McFarland and Bosser [31] distinguish among embedded, implicit and explicit utility functions (goal functions). Spier and McFarland [39,40] tested these basic alternatives in detailed and exhaustive simulation studies. They found that decision strategies based upon explicit utility formulation did not perform as well as those
Fig. 3. Slug availability graph (Glover).
204
J. Greenman et al. / Mechatronics 13 (2003) 195–228
based upon implicit utility maximisation. The reason seems to be that explicit formulations assume a smooth and homogeneous resource distribution in the environment, which originates in the mathematical formulation used to stimulate their design. Decision strategies in which the optimality considerations are mathematically implicit, and are instantiated by simple behavioural rules, not only perform better in simulation studies, but also there is some evidence that they are used by animals [31,34,37]. For these reasons this will be the approach taken in our implementation of decision strategies in the SlugBot. 2.7. Implementation Quadratic trade-offs can be accomplished by simple behavioural rules (see above). Spier and McFarland [39] expanded this approach to encompass the following functional variables: (1) Deficit (d). A vector representing the fundamental state variables of the robot. In the present case, these are E and M. (2) Resource availability (r). A vector representing the rate of change of state that would occur if the robot performed a particular activity. r represents the availability of those environmental resources that the robot needs to alter its deficit. In the present case these would be the availability of energy at the charging station (in terms of the change of deficit per unit time that results from activities associated with recharging), and the availability of resources relating to M (in terms of the change of deficit per unit time that results from activities associated with working). (3) Resource accessibility (k). A vector representing the ability of the agent to exploit r (in terms of the effect on the deficit of a given rk combination). Thus, although slugs may be readily available (high r), the SlugBot may be unskilled at handling them (low k), and the rate of change of hunger (deficit) that results from a particular activity (foraging for slugs) will depend upon a relevant aspect of the environment (r), and a relevant aspect of the animal (k). In the present case, an element of r might be the availability of slugs, and an element of k might be the time it takes for the robot to detect and capture a slug (the handling time). So far, these variables have been regarded as purely functional variables, but one could argue that a robot architecture should be closely related to the functional requirements. Accordingly, we can look at each of these variables in terms of the robot mechanism: (a) Deficit. All that is required to establish the deficit as part of the mechanism is that the robot should be able to measure the elements of the deficit. In the present case the robot must be able to measure E and M. (b) Resource availability. Here the robot must be able to detect the relevant resources. The detected variables are called cues. Thus the robot might have a mechanism for detecting the availability of slugs. This mechanism would yield a value, called the cue strength, which would play a role in the decision-making mechanism. (c) Resource accessibility. Here the decision-making mechanism must have information about some relevant aspect of the robot’s own body that is implicated in
J. Greenman et al. / Mechatronics 13 (2003) 195–228
205
obtaining the resource. Thus if the robot always takes s seconds to detect and pick up a slug, then this value becomes part of the relevant k value. The elements of the K matrix can be thought of as tools. Tools are a means of accomplishing tasks, and a tool may be a physical device, a plan, a behaviour, etc. [31]. Details of the algorithms required to appropriately combine d, r and k can be found in Spier and McFarland [39,41], and will not be given here. Functional analysis has led to the proposal of a particular architecture for our foraging robot. The architecture is appropriate for situations where the costs are likely to be quadratic. Longevity tests of this architecture against alternatives indicate that it is probably the best available for the type of environment in which the robot is to operate. 2.8. The SlugBot as a predator A successful predator has some built-in information about its prey. This will usually include some aspects of prey recognition plus cues to prey availability and prey accessibility. We can see from Kelly and Melhuish [21] how prey recognition is achieved in the SlugBot. Let us now look at prey availability. 2.9. Slug availability The three main factors that affect slug availability are the type of vegetation [12], predation [12], the weather and the time of day. Over 90% of agricultural slugs in the UK are beneath the surface, where they are protected from dessication and predation. Slugs move onto the surface to feed at night, provided the weather is not too dry. The numbers on the surface usually increase rapidly at dusk (see Fig. 3) and may stay high if the night is dark and wet. Normally, there is a peak of surface activity at sunset and dawn, and then the numbers decline to low daytime levels [12]. A predator must be ‘tuned in’ to its environment and behaviour of its prey. In this study the prey species is not permanently depleted and therefore the robot can profitably revisit the same site the next dusk or dawn. On this basis, we can draw up an opportunity profile (see Fig. 4) representing the availability of slugs as a function of time. The opportunity profile indicates the number of slugs that the robot predator would obtain if it foraged efficiently within each successive time slot. Another aspect of prey availability is resource depletion. The SlugBot can, while stationary, remove all slugs within reach of its 2 m arm, a 1.86 m diameter circle. When this patch has been depleted, the robot must move to another patch (or it may decide to move before depletion is complete). Once all the slugs have been removed from a particular area, how long does it take them to reinvade the vacant patch? This is an important question, because we need to know (or the robot needs to know) when the SlugBot should revisit a patch that it has depleted. To answer this question an experiment was setup at the Long Ashton Research Station, an agricultural research station which specialises in crop protection. The work was carried out by Glover [12].
206
J. Greenman et al. / Mechatronics 13 (2003) 195–228
Fig. 4. Opportunity profile.
The main pest of cereals is the field slug D. reticulatum, which eats the crop seedlings. Glover looked at the distribution of slugs within an experimental field, and at the reinvasion of cleared areas by slugs from beneath the surface, and from the surrounding area. He setup a series of plots, half of which had all slugs removed from the surface at around sunset, while the other half had no slugs removed. Soil samples were taken to assess the number of slugs beneath the surface. A soil– moisture analysis was also carried out on each of the 30 plots. In the main experiment 16 of the 30 plots contained circular sampling areas for the study of the reinvasion processes. Eight of these 2 m diameter circular areas were surrounded by barriers to reinvasion (see Fig. 5), the remaining eight being marked out without presenting any barrier to slug movement. Four treatment combinations were used, each being replicated four times. These were: (1) No barrier to reinvasion, slugs not removed from area (NBNR). (2) No barrier to reinvasion, slugs were removed from area (NBR). (3) Barrier present, slugs not removed (BNR). (4) Barrier present, slugs were removed (BR).
Fig. 5. Slug barrier (Glover).
J. Greenman et al. / Mechatronics 13 (2003) 195–228
207
Fig. 6. Reinvasion graph.
Slug counts were conducted shortly after sunset. Areas were taken in a random order, so that specific plots were not visited at the same time of night. Counting was conducted in a systematic manner, starting at the perimeter of the plot, so as not to count slugs that may move in during the counting process. When slugs were removed, their position in the sampling area was recorded. Overall, the results of this study revealed three main points: (1) Removal of slugs from an area on one night had no effect upon the surface slug density the next night. (2) The presence of a barrier to reinvasion of a vacant area also had no effect. (3) The density of slugs in plots with barriers was not affected by repeated removal over six weeks (Fig. 6). It seems that vacant areas are repopulated by slugs coming up from below the surface, and that there is a carrying capacity effect, which limits the number of slugs on the surface. This limit may change according to the weather as slugs are usually more numerous on the surface when the soil is damp. From the robot’s point of view, if it revisits a patch that it has cleared on a previous night, the slug availability will be unaffected. Moreover, a robot could probably inhabit an area for a number of weeks without exhausting the slug supply. This finding is important, because it means that it is advantageous for the robots to be territorial.
2.10. Slug accessibility There is a maximum rate at which the SlugBot can harvest slugs. Once a slug is detected, the robot must open its grabber, pick up the slug, retract its arm, deposit the slug, extend its arm and reinitiate the scanning routine. Even when slugs are very plentiful, the robot cannot complete this cycle in less than 10 s.
208
J. Greenman et al. / Mechatronics 13 (2003) 195–228
D. reticulatum vary in size from a few milligrams to 700 mg. And it is obviously not worthwhile for the SlugBot to capture the very small ones. The question of what size of slug it would be profitable to gather is a matter for further research. For the time being we assume that only slugs over 500 mg are of interest. If the maximum possible capture rate is one slug every 10 s, then there will be a maximum corresponding rate of energy intake. This will depend upon the energy-value of each slug to the robot, and this, in turn, will depend upon the energy losses involved in foraging for slugs, and the conversion efficiency of slugs into usable energy. Within a given foraging session, the accessibility (k) will have a fixed value. It could change in the long term, if the robot could learn to improve its slug-catching skills (see [41]). The slug accessibility could also change as a result of environmental changes. For example, it would be lower if the vegetation were such that the slugs were harder to detect or pick up. 2.11. Quantitative aspects The basic cycle, however defined, consists of work – find station – refuel. We can now define a unit cycle as the time from the end of one full refuel to another. A full refuel implies a certain amount of necessary time at the refuelling station, a certain amount of time spent working, and a certain amount of time spent finding the station again. The time taken to perform these tasks will depend upon the exact circumstances, and the advantage of defining the unit cycle is that it can be tailored to individual real robots, that vary in weight, battery capacity, rate of energy expenditure, etc. Each robot has a characteristic unit cycle that can be determined empirically. In the present case, the unit cycle for the SlugBot is probably in the region of 16 h. The fully charged battery will allow for about 3 h foraging time, and it will take about 12 h to recharge the battery. However, slug availability varies on a 24-h cycle (see above) so the SlugBot will have to fill in the nine extra hours by resting, either at the recharging station or in the field. When resting we estimate that the robot will consume only 45 J/h. The robot remains stationary, with all power-consuming mechanisms switched off, except for its internal clock. This switches on some sensors, say every 15 min, so that the robot can periodically monitor its environment. These sensors will include temperature, humidity and light receptors, which give an indication of slug availability. If it is dark, warm and wet, even in the daytime, then slugs are likely to be active on the surface. While when foraging, the robot will consume about 5460 J/h. This includes scanning, catching, and moving to the next patch. The SlugBot is able to scan a 1.86 m diameter patch of ground while stationary on the basis (see [12,38]) of 108 large surface slugs per patch (in dry conditions). This is approximately 10 slugs/m2 . We estimate (from laboratory tests) that the SlugBot will take 1100 s to clear a patch, and 13 s to move to the next adjacent patch. Whilst foraging the robot will expend about 130 kJ/h, this figure including scanning, catching, and moving to the next patch. On average the robot ought to be able to catch at least 1048 slugs in 3 h, expending 390 kJ. Since the battery capacity is 432 kJ, this leaves 32 kJ to get home. Each slug captured weights at least 500 mg and is worth about 1.5 kJ gross [38]. This
J. Greenman et al. / Mechatronics 13 (2003) 195–228
209
figure is arrived at by discounting 85% of the mass as water and using only 75 mg of the usable biomass. Obviously, the amount of energy expended in returning to the recharging station depends upon how far away the robot is when it decides to return home. Based upon our laboratory estimates so far to travel 1.8 m will take 13 s, and expends 1 kJ. Each patch is 1.8 m in diameter, so if the robot covered 10 patches in 3 h of foraging, the maximum distance home would be 18.6 m, and this would take it just over 2 min, and consume 10.3 kJ. So the total energy expenditure per day (assuming one 3 h foraging trip) is in the region of 400 kJ, and the gross energy gained from foraging is 1572 kJ. This means that the efficiency of the conversion process of slug biomass into usable electricity must be in the region of 25%, for the SlugBot to break even in terms of energy balance. However, McFarland [30] has shown that by optimising the robot’s behaviour it is possible to halve this figure.
3. Generating fuel Our robot runs on electrical power. It catches slugs, and these somehow have to be converted into electrical energy. One possibility is to dry the slugs and then combust them, producing energy in the form of heat. The problem here is that heat conversion to electricity generation is an inefficient process. An alternative is to digest the slugs and ferment the digestive product. This produces an energy output in the form of fuel-chemicals or gases. 3.1. Microbial digestion of organic matter This occurs naturally in the intestinal tract of animals including insects, amphibia, fish and mammals including ruminants and humans [25]. The microbiology is very complex and each animal species will have a diverse but distinct mix of organisms that have grown and adapted to the environmental conditions existing within their intestinal tract. However, despite the microbial complexity, the digestion process consists of three important stages: (a) hydrolytic fermentation, (b) acetogenesis and (c) methanogenesis. These stages are also evident in sewage digestion or in any vegetable/animal rotting process such as composting. The particular ecological mix of microbes in any example will depend mainly on the physico-chemical environment including the temperature, the pH, the degree of aeration (oxygen tension) and the nature of the substrates present. These dictate the main types of microbes that will flourish and in turn dictate the overall nature of the end-products. In animals, further refinements to their gut ecology may occur due to the secretion of host chemicals (e.g., bile salts) which selectively inhibit certain species over others. 3.2. Hydrolytic fermentation Organic matter is composed of polymeric materials (proteins, lipids, and polysaccharides) which are hydrolysed by microbial hydrolytic enzymes into
210
J. Greenman et al. / Mechatronics 13 (2003) 195–228
corresponding monomers (amino acids, glycerol/fatty acids and sugars, respectively). In the first step, species such as Bacteroides, Clostridium, Eubacterium, Peptococcus, Peptostreptococcus, Propionibacterium and Lactobacillus ferment the substrates to produce short-chain fatty acids (mainly acetate, propionate, butyrate, lactate and succinate) plus smaller amounts of alcohols and esters. In addition, carbon dioxide and hydrogen gases are evolved. The microbes involved in these processes can be termed the hydrolytic fermenters.
3.3. Acetogenesis The second step of the process involves a group of microbes collectively known as the acetogenic bacteria. These include Acetobacterium, Peptococcus, Propionibacterium, Syntrophobacter and Syntrophomonas which are able to further reduce the fatty acids to give acetate, hydrogen and carbon dioxide.
3.4. Methanogenesis The third step of the digestion process involves two different routes to produce methane. One of these routes involves the acetogenic methanogens (Methanosarcina, Methanospirillum, Methanothrix) which convert acetate to methane plus carbon dioxide. The second of these routes involves the hydrogenotrophic methanogens (Methanobacterium, Methanobrevibacterium, Methanomicrobium, Methanococcus, Methanogenium), which are capable of utilising the hydrogen and carbon dioxide to make methane plus water. The balance between the two groups of methanogenic bacteria has to be maintained. If the hydrogen concentration gets too high (and ‘surges’ may occur following slight changes in environmental conditions such as temperature, pH or type of nutrient feed) then it may exceed the capacity of the slow-growing methanogens to utilise it. A high partial pressure of hydrogen can change the metabolic pattern of the acetogens whereby they produce less acetate and more of the higher-chain fatty acids and lactate. These products cannot be utilised by the methanogens. The acids buildup, the pH drops and the whole methanogenic process becomes ‘stuck’. The microbes involved in the first parts of the process have relatively fast growth rates (mgt 0.5 to 10 h). The acetogens have slower growth rates (mgt 10 h to 5 days) but the methanogens are very slow growing (mgt 10 h to 2 weeks). The growth rates are very dependent on the temperature of the whole process. It is the slow growth of the methanogens that represents the rate-limiting factor in the whole process. To initiate from scratch, a digestor may take weeks to get going. If a large inoculum of activated sewage sludge was used, then the process can be initiated more quickly. A gram of organic matter (dry weight) can yield up to 60 l of methane a day. The production and utilisation of fuel acids (acetic, propionic and butyric) could be a more rapid process with yields of up to 0.6 g of acid per gram of substrate [8].
J. Greenman et al. / Mechatronics 13 (2003) 195–228
211
3.5. Theoretical energy from organic matter Organic matter is composed of polymeric materials (proteins, lipids, and polysaccharides). If sufficiently dry they can be combusted to produce energy in the form of heat. The total amount of energy they contain (assuming complete oxidation by oxygen in air; as determined by bomb calorimetry) would be between 4.1 and 9.3 kcal/g ( ¼ 17–39 kJ/g), depending on their precise composition. Polysaccharides (e.g., cellulose, starch, pectin) are built up of repeating hexose sugar units ðC6 H12 O6 Þ and produce less energy than lipids which being less oxidised to start with (their empirical formula approximates to C18 H36 O2 ) produce higher heat values. The calorific value of protein is intermediate, containing about 5–6 kcal/g (21–25 kJ/g). Similar amounts of energy can also be extracted through metabolic ‘oxidation’ by living cells. This assumes that the organic matter is in a form that can be utilised efficiently and that the cell species in question uses respiration, a mechanism that uses oxygen as the end terminal electron acceptor. Estimations by Wilkinson [50] suggest that foliage contains about 4% carbohydrate by weight which amounts to 0.82 kJ/ml for liquefied matter (at 4% w/v). According to Wilkinson [50] if food energy were to be converted to an electrical form this would yield 5 kWh/kg for a pure hexose sugar, or 0.2 kWh per litre of liquefied vegetable matter (at 4% w/v). Wilkinson [50] comments that ‘‘This result is surprisingly similar to the energy density of a Lithium-ion battery, with the considerable difference that the chemical energy in food is not readily available as electricity’’. Anaerobic digestion can therefore convert feedstock into a mixture of highly reduced compounds such as fuel chemicals (acetate, butyrate, ethanol) or gases (e.g., hydrogen, methane) which can be converted into energy by other human technological means to produce electrical power. The following strategies are available: 1. Volatile fuel molecules (methane and hydrogen) are combusted by ‘burning’ to produce heat energy that in turn can power a conventional generator. (In essence, what a conventional anaerobic digestion sewage works does.) This process is highly inefficient on the small scale. 2. Catalytic combustion of the fuel chemicals can be achieved with conventional fuel cells. This is very much the same as the first strategy above with the exception that the fuel cell is more efficient and may be capable of using the fuel molecules at concentrations below those required for combustion. However, with the exception of methane and hydrogen, energy will be needed to concentrate and/or volatilise the fatty acids or alcohol. 3. Digestion of substrates to produce microbial reducing power which can be extracted using a microbial fuel cell (see below). 4. We can employ an integrated device that combines the above strategies (particularly the last two). 3.6. Microbial fuel cells Microbial fuel cells [1,45,51] are still in their early days of development but could be coupled to anaerobic digesters to convert microbial reducing power directly into
212
J. Greenman et al. / Mechatronics 13 (2003) 195–228
electrical power. In anaerobic conditions, fermentative organisms typically acquire their energy by glycolysis where cell permeable substrates (e.g., glucose) are transformed by enzymes, ADP and NADþ into pyruvate, ATP and NADH. (NADþ ¼ nicotinamide adenine dinucleotide). ATP and NADH have been described as the ‘energy currency’ of the cell; ATP is energy-rich and can be used to drive the biosynthetic (anabolic) reactions in the cell whilst NADH is a source of electrons (reducing power). In order for glycolysis to continue, NADH must lose its electrons and be reconverted into NADþ . In anaerobes, this oxidation is carried out by the cells transforming pyruvate into more reduced end-products, including acetate, propionate, butyrate, formate, lactate, hydrogen and ethanol, depending on the types of species present in the microbial consortium. The majority of these transformations use dehydrogenase reactions to reduce the pyruvate into fuel compounds. However, they also utilise NADH and regenerate NADþ . Microbial fuel cells use a chemical mediator to scavenge the electrons away from NADH. The microbial fuel cell consists of the anodic chamber containing actively metabolising cells, the cellpermeable electron-mediator and an electrode which may typically be made from carbon. This side of the half-cell must be kept anaerobic. The other half-cell is kept in an oxidised state, using air to oxygenate the cathode electrode; in prototypes this electrode is made out of finely divided platinum. The two half-cells are in contact through a special ion-transfer membrane (e.g., fluorocarbon). Thus, electrons from NADH within the microbes are donated to the anode via the electron mediator. Protons (Hþ ) (also a product of microbial fermentation) diffuse from the anode chamber across the membrane where they combine with oxygen in the cathode halfcell. The cathode supplies the electrons required for this reaction. The result is a current flow from anode to cathode, which in some prototypes may be as high as 1 W per litre of analyte, or 2 A continuously at a typical voltage of 0.5 V per fuel cell [50], at their present state of development, MFCs are not suitable for ‘open’ systems the chemical mediators are lost by dilution. These would have to be replenished (or recycled?). Chemical mediators that have been used in prototypes include thionine, methylene blue and hydroxynapthoquinone [2,7,24,46].
3.7. Design considerations for a robotic digester Microbial systems require that the microbial cell can take up substrate molecules. This requires hydrolysis of the macromolecular (polymeric) feedstock. Physical breakdown of the feedstock (analogous to mastication in an animal or homogenisation or milling in an industrial process) is also a necessary prerequisite to reduce the particle size of the food, increase its total interfacial area and increase the rate of the breakdown process. It is at this stage in real digestion (e.g., in a mammal such as a human being) that digestive enzymes and secretions (saliva, gastric juices, bile salts) are added to start the hydrolytic de-polymerisation process. For some mammals (e.g., ruminants) the nature of the feedstock (grass, mainly cellulose) is such that microorganisms are essential for the breakdown of the macromolecules. Mammals have not evolved the cellulase enzymes necessary for attacking cellulose. For this
J. Greenman et al. / Mechatronics 13 (2003) 195–228
213
purpose, ruminants have evolved a special stomach (the rumen) which allows cellulolytic microorganisms to flourish. It should be noted that microbes produce a wide repertoire of hydrolytic enzymes which means that even in the absence of host mammalian enzymes, they are capable of hydrolytic breakdown of all the common macromolecular substrates; thus digestive aids in the form of enzyme supplements may not be necessary for a robotic digestor. Artificial anaerobic digesters were originally developed to run as a batch process. Here, feedstock is mixed with microbes in a large vat; nothing goes in and apart from methane and other gases, little comes out until the fermentation is complete. In nature (e.g., in ruminants) this does not happen; the system is continuous. Like a conveyer belt approach, food goes in at the front end, biotransformation and chemical energy abstraction occur somewhere in the middle and waste is removed at the far end. In microbiology this is analogous to continuous culture and in both cases the systems are operating as open systems where inputs and outputs reach an approximate equilibrium described as a dynamic steady-state. It should be noted that microorganisms in their natural environments invariably grow as mixed species biofilms [5]. Biofilms consist of layers of microbial cells stuck to surfaces (e.g., adherence of cells to the inner surfaces of the intestinal tract, which in many animals is specially adapted to promote this behaviour). Biofilms have many advantages over their planktonic (unattached suspended cell) counterparts. This includes a greater stability to changes in environmental conditions (improved homeostasis) and colonisation resistance (competitive exclusion) against ‘invading’ microbes from elsewhere (e.g., entering with the feedstock). Microbial growth in biofilms ensures that large ratios of cells are retained to continue the digestion process as the contents move along. The high ratio of microbial cells to substrate molecules helps to maximise the biotransformation rates. Another important process in the gastrointestinal tract of mammals is the selective uptake of ‘fuel’ molecules. In a ruminant such as a cow the fuel molecules are mainly acetate, butyrate and hexose sugars which are adsorbed into the bloodstream. Many animals also have methods for water readsorption in the large intestine that promotes electrolytic balance, conserves water for the animal and increases the concentrations of microbial cells and substrates thus promoting biotransformations. Incalcitrant biomass material is finally excreted as semi-solid waste matter. Further energy abstraction would require incineration. 3.8. Nature of the diet To convert organic matter into fuel chemicals requires the correct balanced diet of substrates. Too much protein in the diet results in too much hydrogen sulphide in the final output. This has consequential bad effects, both in mammals (e.g., ulcerative colitis in humans; [9] and for fuel cells which are inhibited by sulphide. Sulphide production can also inhibit methanogenesis. However, sulphide anion can be removed from solution by electrocatalytic means which might remove the need for sulphide gas scrubbing and its attendant practical difficulties (e.g., replenishment of chemicals or charcoal filter material). A feedstock that is low in sulphur may also mitigate against sulphide production. Feedstock devoid of nitrogenous compounds
214
J. Greenman et al. / Mechatronics 13 (2003) 195–228
can result in lower than optimal yields of cells or products or slower biotransformation rates. A diet of slightly excess carbohydrate can give rise to a microbial phenomenon known as ‘energy spill’ where excess reducing power is dissipated from the cell [36]. It is this power that may be most usefully ‘tapped’ in the design of a microbial fuel cell. 3.9. Physical factors to take into account in artificial digestion When the size increases, the ratio between surface area and volume or weight decreases. This simple fact determines the limits of size for bioautonomous robots. On the small size, mass–energy transfer becomes easy in relation to bulk but heat loss and rapid changes of conditions (e.g., air diffusion into what should be an anaerobic process) may become a problem. On the large scale, the larger and heavier the robot is, the larger the energy and fuel demand. A large digester is more difficult to mix and may in itself be the most significant weight contributor to the total weight of the robot. Environmental homeostasis (temperature and anaerobiosis) is easier to maintain on the large scale but selective removal of fuels by membrane technology (e.g., electrodiffusion) which relies on surface area may become problematic. For an open system, the dilution rate (D) is given by dividing the flow rate (f) by the volume (v). If the dilution rate is greater than the rate of the growth of the key groups of organisms, then culture washout will occur with subsequent slowing down of the rate of product formation. If the dilution rate is too slow, the rate of energy input to the robot is too low to support demands. Typically, the dilution rate for mammals is of the order of 0.01 to 0:1 h1 (equivalent to a mean generation time of 60 to 6 h). The majority of anaerobic digesters seem to work quite well at temperatures of 35 °C at neutral pH (7.0). This is close to the conditions occurring in the large intestine of many mammals. The use of high temperature evolved microbes (thermophilic species) could have important benefits with regard to speeding up the rate of the process. For example, thermophiles operating at a temperature around 75 °C could have rates of transformation over 10-fold greater than microbes operating at 35 °C. One species of thermophilic methanogen (Methanopyrus kandleri) isolated from marine hydrothermal vents can grow at temperatures above the boiling point of water. In an artificial digester, the stages of anaerobic digestion could be teased apart to correspond to different units or cassettes in the system. One could envisage a future gastrobot which has a tubular digestive system split into units or cassettes that carry out different functions depending on the needs of the systems. The cassettes would be inoculated with different communities of microflora which would establish as biofilms within the cassette vessels. For forage on high cellulosic material, a cassette analogous to the rumen would contain species with cellulase activity capable of digesting and fermenting the material into acids. A different cassette containing proteolytic, amylolytic or pectinolytic activities could be chosen for protein-rich, starch-rich or pectin-rich substrates, respectively. Another cassette could be supplied to ensure that the right microflora and environmental conditions were occurring for acetogenesis whilst a final cassette would contain methanogens and could be
J. Greenman et al. / Mechatronics 13 (2003) 195–228
215
controlled and optimised (pH, temperature) so as to efficiently produce methane. For any given flow rate through the whole system, the sizes of the cassettes could dictate the mean residence time of the contents. A larger cassette may therefore be required for the slower processes (analogous to the small intestine leading into the large intestine in man and animals). Mixing and flow could be maintained by a peristaltic pump. Provision for selective uptake of acids (e.g., by electroosmosis) or phase exchange of gases would also be needed (further cassettes?). For a human sized robot it might be expected that an artificial ‘gut’ would equate in size to the natural system. For humans this represents a tube approximately 10 m in length. The colon is about 150 cm in length with a total volume of about 0.5 l [6]. In the UK the mean transit time of gut contents is 70 h with an average stool weight of about 220 g. About half of this consists of gut bacteria [43]. In a future system ideal monitoring/control of the process could be obtained by measuring and controlling the temperature, pH and concentrations of important products (fatty acids and methane). Monitoring would use sensors and biosensors coupled to a microprocessor which by triggering actuators or small pumps could release reagents to neutralise the pH or otherwise correct the physico-chemical conditions. 3.10. Energy extraction Once at the charging station the robot will transfer its load of slugs into a fermentation hopper. Here the slugs will be anaerobically fermented to produce biogas, which is a mixture consisting mainly of carbon dioxide and methane. The biogas will then be passed through a specially developed tubular solid oxide fuel cell that directly produces electricity. Modern fuel cells of this type can produce electricity from biogas containing methane at concentrations of 20% or less [42]. Under development at the University of South Florida is the Gastrobot robot that uses microbial fuel cells to produce electricity from a solution of pure glucose [49]. Unlike the system we propose, their anode is directly placed into the organic food source. It is claimed that with fuel cells of this type, efficiencies of up to 80% [49] can be expected in the conversion of available food energy to usable electricity. However, at the moment, the exchange membrane in such designs would be ‘clogged up’ by ‘real’ food sources such as slugs. The purpose of a fuel cell is to convert the chemical energy of an organic molecule into electrons. In biological cells this is done across a lipid membrane, using aqueous solutions and enzyme catalysts. The man-made equivalent is the hydrogen-powered fuel cell, in which the preferred membrane is a perfluorinated sulphonic acid polymer with platinum catalysed carbon electrodes, separating hydrogen and oxygen in the gas [4]. With hydrocarbon fuels this membrane does not operate effectively and ceramic membranes are preferred, allowing the fuel to react at higher temperatures, again in gas phase conditions. A typical high temperature fuel cell based on a zirconium oxide membrane is shown in the diagram. It consists of a white ceramic tube with a thin wall which acts as an oxygen ion conductor [22]. On the inside is a fuel electrode, usually a mixture
216
J. Greenman et al. / Mechatronics 13 (2003) 195–228
of zirconia particles and nickel particles, with a nickel wire to collect the electrons. Outside the cell is the oxygen electrode, a black conducting oxide material, with a
Fig. 7. Zirconia fuel cell.
Fig. 8. Fuel cell system.
J. Greenman et al. / Mechatronics 13 (2003) 195–228
217
wire wrapped around it supplying electrons. Fuel is flowed through the tube while air circulates on the outside. This provides a potential of around 1.1 V with a typical 2 power output of 0:2 W=cm at 800 °C (see Fig. 7). The principle of operation of this device is electrochemical. Oxygen is reduced to O2 ions at the cathode and the ions are transported through the membrane to release electrons at the anode, where fuel is oxidised. The electrons then do work as they circulate around an external circuit. The benefits are high efficiency and total conversion to water and carbon dioxide. A 1 kWe device could operate at 50% efficiency with low maintenance cost. Heat is the other product, which is used to maintain the temperature of the device and also to process the raw fuel into more suitable hydrocarbons. A fuel cell system tested at the University of Birmingham is shown in Fig. 8. A fuel such as methanol or formic acid is evaporated in a carrier gas such as carbon dioxide and passed down the red-hot zirconia cell. Oxygen is obtained from the surrounding air in the temperature controlled furnace. Power is then released and measured using a potentiostat, while reaction products are analysed by mass spectroscopy. These two fuels can react easily without carbon formation, as shown in Fig. 9. Longer chain molecules such as ethanol, acetic acid and butyric acid tend to foul up with carbon unless a reforming gas such as carbon dioxide or steam is added to the fuel. Carbon deposition causes the cell to die after a few minutes as demonstrated in Fig. 9 bottom line. Thus, the fuel cell operation is highly dependent on the fuel quality. Small quantities of sulphur, greater than 10 ppm, can also destroy the electrodes. The flow rates in the cells are small. Typically the fuel velocity in the 2 mm diameter tubes is less than 1 cm=s. The operation is near room pressure so there is little
Fig. 9. Power output graph. Power output with time; results on various fuels for a zirconia fuel cell at (800 °C).
218
J. Greenman et al. / Mechatronics 13 (2003) 195–228
stress on the components which can therefore be constructed with lightweight materials. Thus, the zirconia fuel cell offers an ideal opportunity to use biofuels directly in a robot power supply.
4. Construction and design of the SlugBot The robot will have to perform in muddy, wet, fields unprotected from the cold, rain and winds. The engineering emphasis is on energy efficiency, reliability and robustness. The combined environment and constraints make for a severe test of any mechatronic system. Let us consider the main activities involved in slug harvesting. Consider the following scenario: a robot decouples itself from the central digester having charged its battery. Using on-board GPS and low-level reactive sensing it makes its way in the dark, over muddy uneven terrain to the harvesting site. In doing so it must deal with obstacles and hazardous features such as deep furrows. Once at the site it needs to efficiently search for and detect slugs, which will inevitably involve sensing, processing and robot movement. During this harvesting phase the slugs need to be collected and stored, as soon as they are detected (see below). At an appropriate time the robot will cease searching and then return to the digester. Once back at the digester it will dock and start recharging, and then it will need to transfer its load into the digester and then possibly having carried out some maintenance task on the digester (e.g., waste removal). Once the robot has recharged its batteries, it is ready to begin the operational cycle once more. All of these operations consume energy. Energy savings can be achieved in several ways: by constructing the robots using light but strong materials like carbon fibre and aluminium; by using decentralised modern low-power controllers and electronics were possible – instead of a single high-speed central processor – thus allowing currently unused devices to be shut down; and by the use of physical designs and control strategies designed to optimise efficiency. Our prototype ‘SlugBot’ robot is illustrated in Fig. 1. The robot uses a four-wheel differential drive system to move across muddy fields. This drive system has the largest energy budget. A trade-off was made between rough terrain capability vs energy demand. A wheeled system was found to be more energyefficient in agricultural fields than a tracked system. Although tracks offer better grip the crop often becomes caught between the links thus causing damage to the crop as well as increasing energy consumption. In order to minimise ground deformation/ loading and hence to save energy, balloon tyres are used. Currently skid steering is utilised since it is mechanically simpler and hence more robust than Ackermann steering. A single motor drives both the left wheels through a toothed belt arrangement; likewise the right wheels have a similar drive system. Although skid steering allows the SlugBot to turn on the spot this will be avoided through path planning to reduce energy consumption. To further reduce energy consumption, scanning and catching of slugs are carried out by a lightweight arm that can search for slugs within a radius of 1.8 m from the robot. When foraging, the robot moves into the centre of the region to be scanned. Scanning involves moving the arm to the closest position to the main chassis, where
J. Greenman et al. / Mechatronics 13 (2003) 195–228
219
no part of the arm will collide with the robots’ main base. The arms end effector, which consists of the slug detection camera, a gripper and ultrasonic sonar transducers, is then scanned 360° around the robot. At the end of each scan the end effector is moved further away from the robots’ chassis and scanned around the robot again. As soon as a slug is detected it is collected, and deposited into a storage hopper on the robots’ chassis. Scanning will then recommence from the location that the slug was collected. The question of deploying the most efficient scanning and harvesting tactics is currently being investigated. Once the surrounding terrain has been sufficiently harvested the whole robot moves on to a new start point and the scanning routine is resumed once more. Design of the arm involved a trade-off between its length and power consumption; on the one hand the longer the arm the greater the area that can be scanned without moving the whole robot but a longer, hence heavier, arm requires more energy to move. It was calculated that an arm length of about 1.5–2 m would produce the greatest energy saving with the expected slug population densities (10 collectable slugs per square metre), for this reason we have opted for a 1.8 m long arm. To meet the criteria of lightness and stiffness, the arm is made from two sections of tubular carbon-fibre connected by a hinged joint. With the exception of the bearings the rest of the arm is constructed from aluminium for lightness. Both the arm’s control motors are mounted at its base to reduce the arms’ weight and hence, to reduce inertia. The drive to the ‘elbow’ joint is transmitted from the base through a toothed belt arrangement running inside the first tube. Self-locking worm gears allow the arm to be held passively in position. The motors also employ slotted opto-encoders to provide sub-degree positional and velocity feedback. Slugs are mainly available at particular times; about 2 h after sunset and for 2 h before sunrise. This limited time window means that the rate of slug-catching must therefore be as high as possible during these times. The arm/gearbox/drive/controller combination allows the arm to move from full retraction to full extension (or vice versa) in under 1.5 s using trapezoidal closed-loop velocity control. As shown in Fig. 1 the arm is fixed to a powered turntable which allows it to be rotated fully around the robots’ base. The batteries are mounted on the back of the turntable as a counter-balance for the arm; at the moment power is transferred to the main chassis via a pair of cables. In the future these will be replaced by slip rings, with the control signals being sent from the turntable to the main chassis via radio; thus allowing the turntable to be continuously turned in the same direction. The chassis is big enough so that the wheels are mounted in positions that ensure that the robot is stable at all arm extensions. Several design iterations were required to construct a lightweight gripper, which is capable of picking up and releasing wet and dry slugs irrespective of size, orientation and sliminess. It was found that dry slimly slugs are very sticky, making them hard to drop; whilst wet slimly slugs are slippery, making them difficult to pick up. The current gripper design is constructed around a three fingered system, with each finger at 120° to one another; which allows a slug, under 75 mm in length, to be picked up regardless of its orientation. Each of the fingers has a plate, which all close together underneath the slug, thus allowing slippery slugs to be collected. Each finger also has a wiping blade ensuring that the sticky slugs are scraped of the fingers; the slug does
220
J. Greenman et al. / Mechatronics 13 (2003) 195–228
Fig. 10. Gripper.
not get enough time to stick to these wiper blades. Small feet on the ends of the wiper blades combined with a passive gimbal system at the grippers’ ‘wrist’ allow the gripper to passively conform to irregularities in the soil when picking up a slug. The gimbal also allows the gripper and hence the image sensor to be held perpendicular to the surface. It can also be locked to prevent the gripper swinging during high-speed scanning. The gripper is opened and closed, and the gimbal is locked/released by a single miniature motor. The image sensor used for slug detection is mounted in the centre of the gripper in between the three fingers, thus providing direct feedback to the target slugs’ position with respect to the gripper. The gripper is shown in Fig. 10. 4.1. Sensory apparatus Although the agricultural fields that the trials will be undertaken are quite large and empty, obstacles may be present from time to time. For this reason the robots will carry the normal complement of obstacle avoidance sensors, required by any mobile robot. Obstacle detection will be achieved using a combination of ultrasonic
J. Greenman et al. / Mechatronics 13 (2003) 195–228
221
sonar and, as a last resort, bump sensors. In addition, two sets of miniature ultrasonic sonar transceivers will be placed in the gripper: one set will point downwards so that the sensor for slug detection can be kept at the optimal distance from the ground regardless of any irregularities, and a second set will face outwards to detect any obstacles in the path of the gripper. According to agricultural experts [3] largetyred robots are not expected to cause any significant damage to crops which are expected to spring back once the robot has passed over them. Once a full load of slugs have been collected the robots will have to return to the fermentation and recharging station, which will be situated at a fixed location. It is imperative that the robot is always able to locate and return to this station before their battery is exhausted. The irregularity of the robot’s terrain will make wheel slip inevitable, which precludes the use of odometry. Locating the fermentation and recharging station will be achieved by using differential global positioning satellite (DGPS) system (which offers sub-metre resolution), and an active infrared localisation system [15,17] (which allows precise autonomous docking [16]). DGPS will also be used for mapping the locations of grazed areas, so that good patches can be found again. 4.2. Slug detection Slugs are visually detected using a low-powered CMOS camera mounted inside the gripper pointing through the fingers. The details of the detection system are described below. The gripper needs to be maintained at an optimal distance of 15 cm above the soil and must be perpendicular to the surface. This is realised by employing a passive gimbal system at the gripper ‘wrist’. The gimbal can also be locked to prevent an unnecessary swinging. The optimal distance from the surface will be maintained by using ultrasonic sensors as well as inverse kinematic information. Once a suitable slug is identified visual and sensor feedbacks are employed to direct the three-fingered gripper onto the target. When the gripper descends over a slug a set of plates under each finger guarantees passive alignment of the gripper normal to the surface. Even under the benign conditions of sparse crops in seed beds the detection of slugs represents a difficult practical problem. Slugs need to be differentiated from non-slug materials such as leaves, stones and soil lumps. We have considered a number of approaches but have opted for a vision-based system which we believe offers the best combination of size, weight, cost and efficiency. VLSI Vision, produce a monochrome CMOS image sensor that is lightweight, relatively low-powered (<175 mW), of adequate resolution (164 124 pixels), and sensitive (down to 0.1 Lux). It has a digital interface, and the maximum frame rate of 60 Hz enables reasonably high scan speeds for the arm. This image sensor also has an adjustable automatic exposure control, and can calculate the average image intensity of the last frame, and perform pixel level thresholding using an adjustable threshold. Since slugs are mainly active at night, we have developed a simple method of extracting slug images from the background by employing a simple form of filtering. An image is recorded with red light supplied by a ring of LEDs surrounding the
222
J. Greenman et al. / Mechatronics 13 (2003) 195–228
Fig. 11. Differential slug images.
camera mounted in the gripper. The camera is fitted with a matching red filter. Under these conditions the backround material of vegetation and soil appear relatively much darker than the slug body of D. reticulatum. Fig. 11(a) shows a 32 mm long Deroceras, together with some grass, under white light – note the relative brightness of the slug and grass. Fig. 11(b) is the same image (except for some movement of the slug) taken using the red light and filter combination – the grass now appears dark whilst the slug is bright. Fig. 11(c) shows how the image of the slug can be picked out from the background by applying a simple pixel-based threshold function threshold ¼ ðc½average image intensity kÞ; where c and k are constants to the red illuminated image of Fig. 11(b). As an unplanned bonus there is an added benefit, in that the threshold scheme does not detect slugs under 15 mm in length, and so filters out small slugs which in fact cost more in energy to retrieve than they can possibly yield. The final and relatively simple stage is the identification of bright patches which are of the correct size and shape [21]. The images are collected and compressed using a low-powered programmable logic array. The image is then passed to an 8-bit low-power 100 mips Scenix microcontroller for shape recognition. Along with the average image intensity the Scenix microcontroller also calculates the pixel threshold using the above function and this value is used directly by the camera to perform active thresholding at the input stage. Each image is buffered into an 8-bit wide 10 ns static RAM by the PAL, with four thresholded pixels being stored in each location of the SRAM. The slug detection system achieves a success rate of over 90%, at a rate of 15 frames per second, whilst consuming just 1 W of power. 4.3. Low-level control architecture The robot system is quite complex: there are a number of motors (gripper, elbow, shoulder, turntable, drive, steering) which must be precisely controlled, and several sensors (shaft encoders, imaging, obstacle avoidance, scan level, battery level, limit switches etc.) which must be monitored. An added complication is that of the turntable, both power and control signals must be transmitted from the turntable to the main robot base and vice versa, but many wires would become too twisted and eventually break. Power will therefore be transferred by means of slip
J. Greenman et al. / Mechatronics 13 (2003) 195–228
223
rings, and data will be transmitted over a low-bandwidth radio link ð28 Kb=sÞ. The high-bandwidth vision processing system (over 1:2 Mb=s) is located on the turntable where it can be linked to the image sensor by cable running along the arm. The only communications to the robot’s base will be to the drive and steering motors, and the bump sensors. Instead of centralising all of the processing, we have opted for a fairly distributed system with the emphasis on local processing, allowing us to use a low-speed two-wire serial ðI2 CÞ bus to link all of the major sub-systems (see Fig. 12). However, the high-level control of the robot is still handled by a single processor. By using a relatively decentralised approach, the overall system complexity is decreased, and the reliability should be increased. For example, all motion controllers share the same design – a PIC microcontroller handles communication and basic processing, and a Hewlett–Packard HCTL-1100 intelligent motion controller manages the shaft encoders and motor drives; each motor controller has a unique address on the I2 C bus. This distributed approach also allows unused sub-systems to be fully shut down, thus reducing the overall power requirements. Since the arm is capable of moving from fully retracted to fully extended (or vice versa) in under 1.5 s, it is potentially hazardous, and we have had to confront some safety issues. Each closed loop motion controller incorporates a simple fail-safe protection system in the form of an independent watchdog. A separate retriggerable monostable controls the enable input of each motor driver. If a monostable does not receive a logic high to low transition every 20 ms from its PIC microcontroller (due to a crash or loss of communication) the associated motor will be shut down. In addition, both the PIC and the HCTL-1100 will stop the motor if it is driven into one of its limit switches. As final layers of defence, each motion controller board features motor current monitoring, and also has an emergency stop input capable of being activated by the high-level microprocessor as well as a stop command from a remote radio transmitter. 4.4. High-level control architecture As we have seen (above) the important variables in foraging for slugs are the current energy deficit of the robot, and the availability and accessibility of the slugs. How is the robot able to obtain the required information? (i) Deficit. The robot will obtain information about its on-board energy ðEÞ by direct measurement. The robot will obtain information about work done ðMÞ by counting and memorising the prey captured. (It may also remember the conditions obtaining when it was captured.) (ii) Cues to availability. The robot will obtain cues to resource availability indirectly. Cues to the likely availability of prey include: Time of day. Prey are usually more available at certain times of day, and the robot can simply use its internal clock to obtain this information. Microclimate. Prey are more likely to be available when certain microclimatic conditions prevail. The robot needs to be able to measure these conditions.
224 J. Greenman et al. / Mechatronics 13 (2003) 195–228
Fig. 12. Block diagram showing the arrangement of the major control elements of the SlugBot.
J. Greenman et al. / Mechatronics 13 (2003) 195–228
225
Recent experience. Memory of very recent success rates provides a good indicator of prey availability. Past experience. Memory of average success rates at certain places in the past will provide a cue to likely prey availability, especially if it correlates with prey habitat preferences. (iii) Tools influencing accessibility. The main tools at the robots disposal are elements of its behaviour repertoire, e.g., homing, docking, recharging, searching, scanning, catching. Deployment of each of these alters the deficit in a characteristic way. Thus the elements of the deficit, the cues, and the available tools can combine in a number of ways and that combination with the highest score determines the behaviour of the robot. The final score in the behavioural final common path [33] is made up from the value of the relative deficit elements, the cue values pertaining at the time, and the value that the robot has set on each tool element. The latter will depend upon the robot’s past experience, whereas the other two are determined by the circumstances at the time. For example, suppose the robot has plenty of energy, but few slugs. The weather cues are strong (i.e., favour high slug availability), but local microclimate cues are weak. The robot has learned that these cues combined with scanning give a poor result, whereas combined with searching yield a high score. The result is likely to be that the robot leaves the current patch in search of a new patch. When it comes across a more favourable microclimate it will switch to scanning, provided other factors (e.g., on-board energy) have not changed too much in the meantime. When the robot catches a slug, the tools that have led it into this situation are reinforced by learning.
5. Discussion In this paper, we have covered the ground required to design and build a selfsufficient robot that can catch slugs and obtain energy from them. First of all, we have tried to make clear that the notion of behavioural stablity is inherent in the energy autonomy. For a robot to be self-sufficient, it must not only be able to close the loop from energy spent on foraging to energy gained from foraging, but it must do so in a sustainable manner, without incurring irrecoverable debt. Secondly, we have tried to specify what we expect from the robot in economic terms. If a robot is going to make its way in the real world, then it has to earn its living, not only in terms of its own energy, but also in terms of pleasing its owner. Thirdly, we have incorporated the notion of optimal behaviour, borrowed from animal behaviour studies. Real animals have to optimise their use of time and energy, because their competitors do so. The same applies to robots in the real world, which will, in effect, compete with the other ways in which humans attempt to solve the same problems (e.g., other modes of slug control). Just as animals have to be successful in reproductive competition, so robots will have to be successful in market competition.
226
J. Greenman et al. / Mechatronics 13 (2003) 195–228
Fourthly, since the robot is, in effect, a predator, we have paid attention to the nature of the prey. Just as animal predators are designed in relation to particular types of prey, so our robot must be designed in relation to slug morphology and behaviour. Fifthly, we have considered the problems involved in generating fuel from slugs. The best approach seems to be to digest the slugs and then ferment the digestive product to produce fuel-chemicals in the form of gases. To extract energy from these gases we propose to use a particular type of fuel cell, that is capable of producing sufficient electrical energy to charge a battery at a home-station. The robot visits the station to recharge its on-board batteries. Sixthly, we discuss the construction and design of the robot. This involves propulsion over fields, in all types of weather, at all hours of day and night. It involves sophisticated apparatus for detecting and capturing slugs, and transporting them to the digester at the home-station. It involves the sensors necessary for navigation and homing, for obstacle avoidance, for microclimatic search, for slug detection, etc. Finally, we discuss some of the problems associated with the low-level and high-level behaviour control.
References [1] Bennetto HP. Electricity generation by micro-organisms. Biotechnol Education 1991;1:163–8. [2] Benneto HP, Dew ME, Stirling JL, Tanaka K. Rates of reduction of phenothiazine redox dyes by E. Coli. Chem Industry 1981;776–78. [3] Bohan, D. Long-Aston Agricultural Research Station, University of Bristol, England, 2000, personal communication. [4] Buchi, FN, editor. Portable fuel cells. In: Proceedings of International Conference. Luzern, U. Bossel, Switzerland; 1999 [chapter 7]. [5] Costerton JW, Lewandowski Z, DeBeer D, Caldwell D, Korber DR, James G. Minireview the customized microniche. J Bacteriol 1994;176:2137–42. [6] Cummings JH, Banwell JG, Engeyst HN, Coleman N, Segal I, Bersohn D. The amount and composition of bowel contents. Gasteroenterology 1990;98:A408. [7] Delaney GM, Bennetto HP, Mason JR, Roller HD, Stirling JL, Thurston CF. Electron transfer coupling in microbial fuel cells: 2 performance of fuel cells containing selected microorganism mediator-substrate combinations. J Chem Technol Biotechnol; 1984. [8] Englyst HN, Hay S, Macfarlane GT. Polysaccharide breakdown by mixed populations of human faecal bacteria. FEMS Microbiol Ecol 1987;95:163–71. [9] Gibson GR, Macfarlane GT. Mechanisms and clinical consequences of H2 disposal in the human colon. In: Duerden BI et al., editors. Medical and dental aspects of anaerobes. Northwood, Middlesex, UK: Science Reviews; 1995. [10] Glen DM, Milsom NF, Wiltshire CW. Effects of seed-bed conditions on slug numbers and damage to winter wheat in a clay soil. Ann Appl Biol 1998:177–90. [11] Glen DM. Ecology of slugs in cereals in relation to crop damage and control. In: Leather SR, Walters KFA, Mills NJ, et al., editors. Populations, and patterns in ecology, intercept. 1994. p. 163–71. [12] Glover C. Reinvasion of cleared areas by the slug Deroceras reticulatum. Project report for MSc in crop protection. University of Bristol, 1998. [13] Holland O. Towards true autonomy. In: 29th International Symposium on Robotics (ISR98). International Federation of Robotics, Birmingham, UK; 1998. p. 84–7.
J. Greenman et al. / Mechatronics 13 (2003) 195–228
227
[14] Holland O. Design of a SlugBot. In: McFarland D, Holland O. editors. Towards the whole iguana. Cambridge, MA: MIT Press; 2002 [in press]. [15] Kelly ID. The development of shared experience learning in a group of mobile robots. PhD thesis, Department of Cybernetics, University of Reading, UK; 1997. p. 2–12. [16] Kelly ID, Hole M. Robot insect II. Internal report, Department of Cybernetics, University of Reading, UK; 1993. [17] Kelly ID, Keating DA. Flocking by the fusion of sonar and active infrared sensors on physical autonomous mobile robots. Conf. Mechatron 1996;1–4. [18] Kelly ID, Melhuish C, Holland O. The development and energetics of SlugBot, a robot predator. In: EUREL: European Advanced Robotics Systems Masterclass and Conference, vol. 2, 2000 12–14 April; Manchester, UK: The University of Salford. p. 8. [19] Kelly ID, Holland O, Melhuish C. SlugBot: a robotic predator in the natural world. In: The 5th International Symposium on Artificial Life and Robotics (AROB 5th ’00) for Human Welfare and Artificial Liferobotics, 2000 26–28 Jan; Compal Hall, Oita, Japan. p. 470–75. [20] Kelly ID, Holland O, Scull M. and McFarland D. Artificial autonomy in the natural world: building a robot predator. In: The 5th European Conference on Artificial Life. Swiss Federal Institute of Technology, Lausanne (EPFL), Switzerland, September. Berlin: Springer; 1999. p. 289–93. [21] Kelly ID, Melhuish C. A slug detection system for the SlugBot. In: TIMR: 3rd British Conference on Autonomous Mobile Robotics and Autonomous Systems. Manchester University; April 2001. [22] Kendall K, Prica M. 1st European Fuel Cell Forum. U. Bossel, Luzern, Switzerland; 1994. p.163–70. [23] Krebs JR, Davies NB. An introduction to behavioural ecology. 3rd ed. Oxford: Blackwell; 1993. [24] Lithgow AM, Romero L, Sanchez IC, Souto FA, Vega CA. Interception of electron-transport chain in bacteria with hydrophylic redox mediators. J Chem Res 1986:178–9. [25] Macfarlane GT, Cummings JH. The colonic flora fermentation and large bowel digestive function. In: Philips SE, Pemberton JH, Shorter TG, editors. The large intestine: physiology, pathophysiology and disease. New York: Raven Press; 1990. [26] Maturana H, Varela F. Autopoiesis and cognition: the realization of the living. In: Cohen R. Wartofsky W. editors. Boston studies in the philosophy of science, vol. 42. Dordecht: Reidel; 1980. [27] Maturana H, Varela F. The tree of knowledge: the biological roots of human understanding. Boston: Shambhala/New Science Press; 1992. [28] McFarland D. Animal behaviour. 3rd ed. Reading, MA: Addison-Wesley; 1999. p. 580. [29] McFarland D. Autonomy and self-sufficiency in robots. In: Steels L, Brooks R, editors. The artificial life route to artificial intelligence Building situated embodied agents. New Haven: Lawrence Erlbaum Associates; 1995. p. 187–213. [30] McFarland D. Energy autonomy of the slugbot. In: McFarland D, Holland O, editors. Towards the whole iguana. Cambridge, MA: MIT Press; 2002 [in press]. [31] McFarland D, Boesser T. In: Intelligent behavior in animals and robots. Cambridge, MA: MIT Press; 1993. p. 305. [32] McFarland DJ, Houston A. In: Quantitative ethology the state space approach. Pitman Books: London; 1981. p. 204. [33] McFarland DJ, Sibly RM. The behavioural final common path. Philos Trans R Soc (Ser B) 1975;270:265–93. [34] McFarland D, Spier E. Basic cycles utility and opportunism in self-sufficient robots. Robot Autonomous Syst 1997;20:179–90. [35] Orions GH, Pearson NE. On the theory of central place foraging. In: Horn DJ, Mitchel R, Stair GR, editors. Analysis of ecological systems. Columbus, OH: Ohio State University Press; 1979. p. 155–77. [36] Russell JB. Strategies that ruminal bacteria use to handle excess carbohydrate. J Animal Sci 1998;76:1955–63. [37] Sibly RM, McFarland DJ. On the fitness of behaviour sequences. Am Naturalist 1976;110:601–17. [38] South A. Terrestrial slugs: biology, ecology, and control. London: Chapman & Hall; 1992. [39] Spier E, McFarland D. A finer-grained motivational model of behaviour sequencing. In: Maes P, Mataric M, Meyer J, Pollack J, Wilson S, editors. From animals to animats 4. Cambridge, MA: MIT Press/Bradford Books; 1996. p. 255–63.
228
J. Greenman et al. / Mechatronics 13 (2003) 195–228
[40] Spier E, McFarland D. Possibly optimal decision making under self-sufficiency and autonomy. J Theoret Biol 1997;189:317–31. [41] Spier E, McFarland D. Learning to do without cognition. In: Pfeifer R, Blumberg B, Meyer J, Wilson S, editors. Animals and animats 5. Cambridge, MA: MIT Press; 1998. p. 38–47. [42] Staniforth J, Kendall K. Biogas powering a small tubular solid oxide fuel cell. J Power Sources 1998;71:275–7. [43] Stephen AM, Cummings JH. The microbial contribution to human faecal mass. J Med Microbiol 1980;13:45–56. [44] Stephens DW, Krebs JR. Foraging theory. Princeton, NJ: Princeton University Press; 1986. [45] Suzuki S, Karube I, Matsukota H, Ueyama S. Biochemical energy conversion by immobilized whole cells. Annals New York Academy of Sciences; 1983. p. 133–43. [46] Tanaka K, Vega CA, Tamamushi R. Thionine and ferric chelate compounds as coupled mediators in microbial fuel Cells. Bioelectricity Bioenergetics 1983;11:289–97. [47] Grey WW. An imitation of life. Scientific American; May 1950. p. 42–5. [48] Grey WW. The living brain. London: Duckworth; 1953. [49] Wilkinson S. A crude but novel carrot powered gastrobot for middle or high school demonstrations. In: Proceedings of the 7th International Conference on Robotics and Applications; 1999. Paper # 304-024. [50] Wilkinson S. Gastrobots – benefits and challenges of microbial fuel cells in food powered robot applications. J Autonomous Robots 2000;9(2):99–111. [51] Zhang X-C, Halme A. Modelling of a microbial fuel cell process. Biotechnol Lett 1995;17:809–14.