6.2.1. Planning robots

In speaking about robots, in particular about autonomous systems, terms from the field of planning are often used. This use of planning language reaches back even to the early times of the artificial intelligence (AI) movement in the 1970s. The roboticists’ fundamental assumption about planning is going back to a statement decades ago:

“Solving problems, searching, and planning are the means with which instructions […] can be obtained […] out of a large knowledge base […] it is a matter of defining the initial state, the goal, and any known intermediate states […]. In general, a number of paths lead to the goal. […] In planning, a means-end table is first drawn up in order to recognize central decisions” [DEC 97].

The classical application in direct analogy to humans in their life-world is that a robot has to find its way through surroundings unknown to it, for instance, by overcoming or circumventing obstacles in moving forward. Another application is soccer-playing robots, which requires coordinating the action of several players. On the part of the constructors, development toward greater robot autonomy is progressing rapidly. This will further increase the demands on planning which the robots will have to do in order to perform the tasks intended for them and to be able to deal with unforeseen, unprogrammed situations in executing the autonomously established plan:

“In the future, robots will be far more than special machine tools. They will work together autonomously with human beings to do complex tasks and to do defined work without exact specification of the course of action in unknown or changing surroundings” [STE 01].

This 15-year-old expectation has in the meantime largely become reality. For example, autonomous automobiles [MAU 16] are nothing else than autonomous robots that move on wheels, transporting people or goods. Street traffic is a sequence of continuous unexpected events. Drones are flying robots that can independently search for their target in a military area. And some robots are already in use in the field of caregiving.

Aspects of planning have played a central role in the theory of AI and of the realization of autonomous artifacts for decades now [POL 95]. A robot as an autonomous system given the task of finding its way through an unknown environment and of carrying out a set task – for example, a transport operation within a building – is one of the most important applications and tests of its performance. This poses the question of the understanding of planning that applies here and of the relationship between the robot’s planning and human planning. If planning competence is ascribed to robots in a more than metaphoric manner, they are incorporated into a “community of planners” – a step toward a socialization of technology [JOE 01]. The importance of such an attribution becomes apparent when compared to philosophical anthropology, in which the ability to plan and the disposal over the linguistic means for imaging possible futures necessary for planning were seen as elements of humanity’s special position (Sonderstellung) [KAM 73]. Answers to these questions obviously touch upon questions of responsibility – who would be responsible for a robot’s planning and for its consequences?

The quotations given here continue to be up to date despite the rapid technological progress that has been made in some fields. You could say that they were ahead of their time. Newer developments such as evolutionary robotics, organic computing or adaptive ambience [GUT 15] lead to advances and new technical opportunities as well as to new sociotechnical constellations [RAM 07]. The fundamental questions regarding the interface between humans and robots remain, however. Examples of these are questions about the collaboration between humans and robots at work [MON 15], the question of whether robots can evolve [GUT 15] and the question of if and when it will be possible for the activity of robots to replace that of humans [JAN 12]¹.

6.2.2. Planning as special type of acting

Because planning is a specific form of action [HAB 68, GRU 00], the introduction of the concept of planning begins with the following definition of action which contrasts action to behavior [JAN 01]:

• – actions can be attributed to actors (as the cause of the action) from the perspective of an observer;
• – actions can be carried out or omitted, not in the sense of any freedom whatever, but on the basis of reasons;
• – actions can succeed or fail, i.e. there are criteria and conditions for success which are commonly explained in instrumental rationality of means and ends.

The classification of something as action is an interpretation made by external observers or by the actor themself [LEN 93]. In the latter case, it is necessary that the actor dissociates himself from himself in order to be able to make the interpretation. Acting thus is no ontological predicate, but an interpretation of the corresponding situation and an argumentatively explicable attribution. If a truck drives by, we do not say that the truck is acting but that the truck driver is acting, whereby a causal relation between the driver’s action (putting on the brakes) and the perceivable effects of these actions (the truck stops) is assumed. Slipping on ice or a coughing bout would not fulfill the criteria for acting, as a rule, but are “occurrences” [KAM 73] – events which simply happen. A coughing bout can neither succeed nor fail. This is only seemingly trivial because there are situations in which coughing can be chalked up as acting: removing crumbs from the respiratory tract, wanting to attract attention in the concert hall or warning a business partner that he is in danger of making a mistake in negotiations.

This shows clearly that the classification of a phenomenon as acting is done through attribution, which is based on an interpretation of each specific situation and its context [SCH 87, LEN 93]. Two coughs may be phenomenologically identical, can, however, possibly through the interpretation, in the one case be categorized as behavior and occurrence, and in the other case as acting. Inasmuch as interpretations can be controversial, the attribution of the concept of acting can also be challenged in individual cases.

Acting in the above-mentioned sense allows for learning. It can be derived from the criteria given above that acting is seen as capable of being improved, mere behavior, however, is not. Knowledge about relations of means and ends and knowledge for correcting faults can only be related to action, and not to behavior. Acting is subject to conditions for success, out of which a measure of success can be derived, and makes learning possible, because – technically speaking – the gap between the target aimed at and the achievements made so far can be used to inquire about the causes of this difference and to think about and introduce possible improvements.

The concept of acting in the sense mentioned above also allows building relations to the concept of responsibility (Chapter 2). With reference to actions, one can speak of reasons, consequences and responsibilities. Inasmuch as humans describe themselves as being capable of acting, they set themselves in a sociocultural context and define themselves as social beings who can develop and discuss actions, who can choose among alternative options for action, who can carry out the actions and who can – both beforehand and afterward – talk about consequences and responsibility (Chapter 2). The concept of action in differentiation from mere behavior is part of the (modern) human being’s self-constitution. However, there are alternative self-descriptions. Cultures are conceivable which do not distinguish between acting and behavior but subsume all of this under behavior. Consequently, these cultures must omit concepts such as responsibility, guilt, justice and injustice. Recent debates on a purely naturalist image of humans have pointed into this direction [JAN 12]. It therefore has to be asked in the following whether and to what extent an attribution of a capacity for action or a planning competence to robots would go hand-in-hand with human acting and planning, or where conceptual differences remain.

Planning is a matter of active preoccupation with future action for the purpose of consideration and preparation. Planning is an anticipatory reflection on purposes and goals or on action schemes without directly actualizing them: a drafting of future options for action in the sense of a test action [SCH 81, STA 70]. The purpose of planning is the previous drafting, reflection and judgment of the possible options for action to the end of preparing the action. Planning allows realizing purposes only indirectly. Only the plan implementation is supposed to realize the purpose, not the plan as such. Planning is a hypothetical and experimental action in the space of possibilities and alternative options of what could be done. It presents itself:

“[…] as a dramatic testing of different, competing possible directions of acting in the imagination […]. Experimentally, various elements of habits and drives are combined with one another, in order to find out how the resulting action would look in case it was initiated” [DEW 22, p. 190].

Schütz [SCH 71, SCH 81] emphasizes, footing on Dewey [DEW 22], planning from the perspective of a phenomenologist and likewise points to the difference between the projected and the real action:

“[…] devising action takes place in principle independently of all real action. Any drafting of action is much rather an imagination of action, i.e., an imagination of spontaneous activity, but is not the spontaneous activity itself” [SCH 81, p. 77].

Planning is devising and preparing goal systems or action structures which prima facie are neither known nor evident: if a decision maker had an action routine already at hand, he or she would not need to plan at all. Planning is an intellectual anticipation of future action and a method for attaining adequate action anticipations [STA 70]. Planning always is only concerned with situations where there is a need for design, preparation, construction, composition, choice and decision making.

An essential attribute of the concept of planning is its second-order purposive rational character [GRU 00]. First, each of the individual action steps of a plan has to be purposive rational and should expectably lead to the realization of certain subgoals. Second, these elements also have to be arranged in a purposive rational manner. The composition of the singular elements has to be done so that the objective as a whole can be attained: planning consists of the purposive rational composition of purposive rational elements [HAB 68]. This second-order purposive rationality implies that planning takes place within the space of reasons and knowledge, and has to be done discursively (see Chapter 4) [GRU 00]. A planning discourse consists of (1) a discourse on the determination of the purposes and goals, (2) the elaboration of alternative options for pathways of how to reach the envisaged goal by certain means and (3) the decision for choosing among the alternative options. This structure will be an important pattern to analyze the planning of robots.

6.2.3. Step 1: Can robots act?

The definition of acting given above leaves open who (or what) comes into question as an actor. It is an empirical question whether the criteria necessary for acting can be fulfilled only by human beings, by certain human beings in certain situations, by rational beings in the sense of Immanuel Kant, by certain animals (e.g. primates) – or even by robots. The definition of acting determined by criteria is open in both directions: not all human beings have to be capable of acting, and actors do not necessarily have to be human. A look at very young children, at demented individuals, at coma patients, at certain types of disabilities and at people with compulsive mental disorders shows that not all human beings can act. Even sleeping humans are not able to act.

Conversely, beings which do not belong to the species Homo sapiens but can nonetheless act are at least conceivable. It is an empirical question of the fulfillment of criteria on the basis of interpretations and reconstructions. Here, it can admittedly come to considerable problems of judgment, when the behavior of a chimpanzee, for example, is supposed to be classified as action. The necessary interpretations could be criticized as mere anthropomorphic misrepresentations because humans do not share their discursive community with primates. The latter also applies to the relationship between human beings and robots. A difference to the primates, however, consists of the fact that the robot, as a construct of human beings, should be better known in its functioning than primates.

The next step now is to ask how the types of robot planning described in section 6.2.1 above appear in the light of the definition of acting given. In short, the three criteria seem to be fulfilled:

• – action causation: there is presumably no question that autonomous robots can cause something, in the sense that the effects of their actions can be causally ascribed to them;
• – identification of success or failure: inasmuch as such robots have a task (e.g. defusing mines, forwarding a message to an address or bringing persons from A to B), success, failure or partial success can easily be determined from the perspective of an external observer;
• – capability to omit: a specific robotic action, for example, circumventing an obstacle, seen from the perspective of an external observer, could have been omitted, analogous to the action of a human agent – namely, if the arguments that were decisive for the action had been different, for instance, due to a different diagnosis of the situation. Inasmuch as human freedom is not to be understood in the sense of a randomizer, but means the freedom to decide on the basis of good reasons, one would have to concede that a robot which chooses one option out of a spectrum of action schemes which is suited to the diagnosis of the situation and to the tasks it has to carry out, would have desisted from this action and have chosen another, if the reasons had been different.

These considerations do not lead to any arguments for denying autonomous robots the capability of acting. However, a consequentialist argument against this conclusion is repeatedly brought forward: the argument of the ascription of responsibility connected with performing an action (see Chapter 2). If the capability to act was assigned to robots, then responsibility would also have to be assigned to them, some say. These voices conclude that because the attribution of responsibility to robots seems to be contra-intuitive, robots cannot be considered capable to act.

A somewhat closer look at the concept of responsibility shows the fallacy in this argumentation. The argument that the attribution of action competence to robots implies the attribution of legal or moral responsibility is based on a confusion of the concept of responsibility (see [GRU 12c] and Chapter 2). The capability to act is a necessary precondition to be made responsible but not a sufficient one. The mere responsibility for causing an action would be applicable to the robot, inasmuch as acting is conceded to it – but this does not automatically imply any legal or moral responsibility of the robot. Also in the case of a causal responsibility on the part of the robot having acted, the attribution of the legal and moral responsibility could lead to the robot’s owner, its operator or its manufacturer [CHR 01, DEC 13].

The argumentation for attributing action competence to robots presented above is subject to a premise, which gives reason for further differentiation. Understanding acting as an attributive term was, in this argumentation, assumed from the perspective of an external observer. Let us now consider a thought experiment: first, an acting human being is observed in this manner. Then, this human being will be replaced by a robot acting in a functionally equivalent manner. If now an external observer would consider this robot, the result of his interpretation should be the same for the robot as for the human being: both are regarded as acting. This thought experiment confirms, on the one hand, the train of thought developed above: robots which replace acting humans act. On the other hand, a difference remains. In the thought experiment, neither the human being nor the robot observed was asked about his/its activities, tasks, diagnoses and reasons. The interpretation was made solely from the observing and reconstructing external perspective. This seems to be artificial concerning human beings: why not asking acting persons for their reasons? In case of the robots, however, asking for their reasons to act might be more difficult or even impossible. This observation indicates a deep-seated difference between humans and robots even in case both are regarded as being capable to act. We will return to this difference later on.

6.2.4. Step 2: What do robots do when they plan?

Usually, two types of planning robots are distinguished according to the differences normally made between artificial intelligence (AI) and artificial life (AL) robots [KIN 97 ]:

• – robots which, based on an analysis of environment data, can choose from a predetermined number of options for action according to a likewise predetermined set of criteria;
• – robots based on neural networks, which can incorporate learning effects and then change the basis for planning and deciding.

In both cases, planning as a preparation for future action obviously plays a crucial role. The first type is of a rather simple nature because the number of actions and set of criteria are predetermined. Planning is, in this case, limited to an assignment of options for action to a diagnosis of the situation. The second type, however, is of particular interest because the action schemes coming into question are possibly created originally by the robot and are not previously programmed as part of a predetermined quantity. Learning through an accumulation of experience from carrying out actions – for example, in moving through an unknown terrain – is at the heart of this type of planning. New ways of acting can result from learning processes in an unforeseeable manner [KIN 97]. This makes a control architecture necessary: the unforeseeable behavior of a robot due to learning could lead to undesirable results. Robots could get out of control. The control architecture has to ensure that the robot’s behavior remains within a defined frame, or that the robot will be turned off.

If planning is understood to be an experimenting test action [SCH 71], the question of the more detailed sequence in a robot’s learning process poses itself. The explanation:

“Learning consists of the reorganization and re-evaluation of the individual links within a neural network […]. We have previously spoken of supervised learning, by which, for example, human beings exercise control. If we go further to unsupervised learning, then we replace the monitoring system by a set of well-defined rules for learning. The entire system optimizes itself according to these learning rules” [SCH 93b]

leads to the conclusion that robot’s learning happens through experience. In reflecting on an autonomous robot of the type AMOS [KNI 94] or ARMAR [ASF 99], it is, first, important that it does not dispose over a prefabricated model of its surroundings but produces one itself through experience and continuously improves and adapts it. An illustrative example of dealing with obstacles is a delivery messenger or courier robot in an administrative body moving around in a building. Through sensor signals, the robot generates a model of its surroundings while it is moving. As long as these surroundings are static, the model produced consisting of walls, doors, elevators, etc., can be used without problems. During operation, the robot constantly checks, by means of sensor technology, whether its model is still up-to-date. If a door, which is normally open, is once closed, it comes to a breakdown of the plan, just as when an obstacle unexpectedly prevents moving ahead. A breakdown in plans shows a deviation between the real situation and the expectations. In such cases, the robot defines the area in which a difference between the model of the surroundings and reality occurs as a region of interest (ROI) [KNI 94, p. 77]. Through experimental handling of the unexpected situation, the robot can gather experience. It can try to bring the obstacle to make way by giving an acoustic signal (the obstacle could be a human being who steps aside after the signal), it could try to push the obstacle aside (maybe it is an empty cardboard box) or the robot could, if nothing else helps, notify its operator. Maneuvers such as parking or making a u-turn in a corridor can be planned in this manner [SCH 95]. One of the most important challenges in this work is classifying the plan breakdowns [KNI 94, p. 80] in order to later diagnose the problem and then take the appropriate measures as fast as possible.

The underlying planning theory paradigm consists of the cybernetic planning model of feedback in a system–environment interaction. The heart of this planning concept [STA 70, CHU 68, CHA 78] consists of a cybernetic feedback loop: a planning system plans to change certain parameters of its surroundings and takes measures to achieve this goal. It then monitors the effects of the implementation of the measures and evaluates them against the expectations. Deviations from the expectations are detected by means of this feedback control mechanism and are taken into consideration in subsequent measures. Learning consists of repeated runs of this cybernetic loop, with a corresponding accumulation of empirical information. A robot’s experimenting with unknown surroundings and the use of the resulting experience can, in fact, be interpreted as planning processes within the framework of cybernetic planning theory.

The reference to the concept of planning introduced above, in particular the specifics of purposive rationality of the first and second order and the necessity of distinguishing and selecting among alternative options (see Chapter 3 and [GRU 00]) do not provide any grounds for rejecting the concept of the planning robot. The robot makes – through sensors – an interpretation of its present situation and compares it with a goal situation. It compiles possible plans of action derived from a knowledge base and then decides on the choice and composition according to a given set of criteria. The specifics of planning, especially the instrumental rationality, are obviously included. Planning-theoretical modeling of the robot’s planning is possible and adequate considering the cybernetic loop.

The robot’s planning in a cybernetic model is, however, an extraordinarily limited type of planning in comparison with the complexity of human planning [GRU 00]. This needs to be elucidated in two directions: (1) by exposing the cybernetic feedback as a poor and deficient planning model and (2) by examining the preplanning agreements:

1. 1) The cybernetic mechanism consists of learning from experience through more or less well-prepared testing and practical trials of action steps in the cybernetic control loop. In the model of adaptive and continuous planning, the robot adapts itself to the conditions of its surroundings. The normativity of planning – namely, to make a plan according to certain aims and, if applicable, to implement it – is neither taken into account in the cybernetic model nor is there a mechanism which could reconstruct this normativity [GRU 12c]. The mechanism of checking the results of planning and comparing the present situation (“what is”) with the desired one (“what ought to be”) simply acts as a substitute for the normativity of the planning goals. The robot is not compelled to determine objectives beforehand, to reason about means for reaching the goals and about possible incidental consequences, but it can try actions and classify the results. What is without doubt sensible in the case of, for instance, AMOS and ARMAR (see above), fails, however, in planning tasks of other types, such as building a house, or of large-scale technical projects where much more elaborated anticipation is needed [SCH 81]. Instead of adapting to environmental conditions, it is, in the latter case, a question of defining objectives and of realizing them by applying appropriate measures. The specifics of this type of singular planning are the normativity involved and the previous modeling and simulation of the entire process including a permanent reflection. In comparison, cybernetic planning is nothing more than an improved method of trial-and-error – a method which, as a rule, plays no great role in normal human planning. Thus, even if robots are capable of planning, although they can only do this within a poor concept of planning.
2. 2) A second type of limitation of robots’ planning competence results from the question on the decisions made before planning. Concrete planning is not free of premises but is based on preliminary decisions through which the space for possibilities, options and searching for the solution of the respective planning task is predetermined. The initial conditions determined before starting the planning process define the stage and set the scene for subsequent planning. They decisively influence the manner in which one can plan and how possible plans could look like. Elements of such preplanning agreements (see section 4.2 and [GRU 00]) include determination of the target areas to be taken into consideration, of the addressed system’s borders, of the range of admitted goals and means and of criteria for choosing a plan among several possible ones. Preplanning agreements are contextual restrictions of the principally conceivable diversity and should reduce contingency. Planning contexts can be distinguished according to whether the preplanning agreements are under the planners’ control or whether they were set for the planners from outside. Robots are, in this respect, in a weak situation.

6.2.5. The difference between planning humans and planning robots

A robot’s planning is, as portrayed, describable in the cybernetic planning model (section 6.2.3). The objectives stated and the target setting are limited: in part, algorithm-like sequences are determined, in part, the knowledge base is predefined, limits are set for the robot through the control architecture and so on. Preplanning agreements (section 6.2.4) have been made which cannot be revised by the planning robot itself. Thus, though the robot’s behavior can definitely be designated as planning, it is a very special and reduced type of planning:

“The behavior of autonomous robots is – by using the techniques of information processing available today – marked by their knowledge base in the form of programs and data, and precognition in whatever form it takes. This knowledge base, its use, and expension is, even in the case of so-called self-learning systems, predetermined by humans in the realization of robot systems” [STE 01].

A first intuition now might be: the planning robot is forced to plan within preplanning agreements made by humans which cannot be changed by the robot – a poor situation. Humans, on the contrary, could be regarded as free to change preplanning agreements. However, the simple juxtaposition of a freely planning human being and of a strongly controlled robot planning falls short. Human planning often also takes place in strongly restricted possibility spaces (e.g. within restrictive employment relationships). It seems, and this would be the resume, that there is a gradual transition from the simple planning of a robot with restrictive planning agreements to free and complex planning processes, which does not necessarily make a qualitative jump at the transition from robot to human being.

In this manner, it becomes possible to reconstruct shifts in limits. Inasmuch as technological progress will increase robots’ planning competence, the previous limits will be shifted. The borders between humans and technical artifacts are becoming blurred. However, Latour’s demand [LAT 87] to speak of robots and human beings in the same language and to acknowledge a complete symmetry between them, tums out to be not very helpful in this connection. Though it is possible to speak of planning robots as well as of planning humans, a complete symmetry between humans and robots cannot be deduced from this as has been shown above. Deriving a full symmetry between humans and robots from the application of the same planning terminology would be possible only by an extreme disregard of the different planning models, the differing disposal over preplanning agreements and the different treatment of the normative level involved. Differentiations have to be made in order to arrive at a better understanding of planning robots and human beings, of their similarities and differences and of shifts between them over time. Only painstaking deliberation and consideration of the differences is instructive: one can, in the comparison of planning robots and planning humans, also learn something about planning humans, namely, the characteristics of human planning and their, in part, very narrow limits, set by heteronomous preplanning determinations, bound by certain terms of reference, e.g. in employment law or in the organization of labor in industrial processes.

When we reflect on planning robots and humans, we also reconstruct ourselves (Joerges [JOE 01, p. 196], with reference to Latour). The fact that we can speak of planning robots does not mean that the planning of humans and robots is to be put on the same level [GRU 12c]. In a certain sense paradox, the use of the same terms for planning robots and human beings intensifies the asymmetry instead of bringing about symmetry. If we reconstruct the work of a messenger robot, we will – formally – find the same action-theoretical structures as when we reconstruct the action of a human messenger as a messenger. The putatively strongly limited objective-setting competence on the part of the robot (in which the tasks are programmed) is no counterargument because, in an environment regulated by employment law, even the human messenger has a very restricted objective-setting competence; in principle, he or she has to do what his or her superior demands, in the scope of his or her job description, to which he or she has consented. On this level, the activities of the human messenger and the messenger robot are equivalent – otherwise, the messenger robot could not replace the human messenger.

Despite this, there is a considerable asymmetry between the messenger robot and the human messenger. A robot which is functionally equivalent to the human messenger, i.e. which brings the same messenger performance, plans the errands and the solution of problems occurring in them in a specific sense and under predetermined initial conditions. The human messenger plans in that he/she fills his/her role, according to an analogous understanding of planning, and with probably similar criteria. While the messenger robot, however, is committed to its role as a messenger through programming and control architecture, the human messenger can abandon this role. The requirement for the ability to omit acting in order to distinguish action from behavior has to be differentiated. For the robot, it is already fulfilled when it has the choice between some few alternative options for action – but yet, it remains within its role. The human messenger, on the other hand, can understand the omission much more radically and can abandon his or her role. For example, the human messenger might follow a strike organized by his or her trade union. Or another example: if a human messenger observes during performing his/her job that a person urgently needs help for health reasons, he/she immediately would stop doing his/her job and instead help the person – the robot would not be able to do this. The measure of the ability to omit planned actions and to move to another track of acting and planning proves to be central to the distinction between planning humans and planning robots, and also to be a parameter for measuring future shifts in this field.