Abstract
When you meet a delivery robot in a narrow street it stops to let you pass. It was built to give you precedence. What happens if you run into a robot that was not trained by or for humans? The existence in our environment of robots which do not abide by human behavioral rules and social systems might sound odd, but is a case we may encounter in the future. In this paper, self-taught robots are artificial embodied agents that, thanks for instance to AI learning techniques, manage to survive in the environment without embracing behavioral or judgment rules given and used by humans. The paper argues that our ontological systems are not suitable to understand and cope with artificial agents. The arguments are speculative rather than empirical, and the goal is to drive attention to new ontological challenges.
Keywords
Introduction
The marriage between robotics, even when enriched with Artificial Intelligence (AI) techniques, and semantic approaches has not worked well in the past. For long time roboticists have been concerned with hardware limitations and search problems in core areas like kinematics control, navigation, object recognition, obstacle avoidance and so on [1,16] At that time, the information stored in the robot used to be carefully selected, and encoded in
Now things have changed. In the last 10 years the continuous improvement of hardware and information processing capabilities has led roboticists to imagine general purpose agents acting in open environments. This vision requires to develop planning techniques for multiple coexisting goals that go beyond traditional robot’s control and navigation, and to build robots that can reason in terms of actions, plans, expectations, other agents’ intentions and possible collaborations. Important technical and conceptual consequences were brought up by this change like, e.g., the distinction between geometric planning and task planning [15], and the distinction between behavior (roughly, how the agent interacts with the environment) and function (that is, how that behavior contributes to the achievement of a goal) [12].
Once the need to enrich the robot with models for environment, goals, actions, functions and behaviors became clear, the community started to investigate suitable semantic approaches, e.g., [13,14]. Today semantic techniques and applied ontology methodologies are largely exploited for a variety of tasks like decision making, belief update, situation assessment, interaction and communication. Interest in ontological modeling is further witnessed by the release of a dedicated standard in the area of robotics and automation [9].
When the aim is to develop general purpose autonomous robots, the information system of the robot must be able to process, integrate, store, recall and update information coming from a variety of sources (e.g., different kinds of sensors as well as different types of collaborators) and deliver dedicated information based on goals, detected environment and decision making processes. Furthermore, the knowledge model, which uses this information to build a view of the environment, must also elaborate possible outcomes, detect the presence of other agents and relevant objects, predict other agents’ goals and future actions. Ideally, such a system takes advantage of techniques in information science and uses ontology as pivot to ensure the reliability of the information management system. This view looks very promising today since important limitations, like memory capacity and processing speed, have been largely removed or attenuated.
What is the role of applied ontology in this setting? Applied ontology has been introduced to overcome interoperability problems, primarily at the semantic level, caused by the existence of different perspectives, e.g., databases developed by different organizations or interpretation mismatches by agents with different roles. What pushed the ontologists to believe in the possibility of information integration was the simple observation that all information is about reality or about human views of reality, since information (via human perception or human designed sensors) as well as its interpretation is human-based and since “reality cannot be (self-) contradictory”, as the motto goes, as long as each agent is locally consistent, everything can be managed to fit. Of course, we have general ontologies that make incompatible choices and are mutually inconsistent. Yet, it is assumed that humans can, at least in principle, understand each other’s ontological system, and even switch from one system to another as needed. Practically, two human agents relying on different ontologies may need to go through an interaction phase to understand each other’s viewpoint but at the end they can correctly interpret the information they exchange. Or so it is believed. How to formally model this from the logical viewpoint remains a problem due, so the assumption goes, to the limitations of today’s formal ontology understanding and of the adopted logical systems.
In other domains one does not need to make these assumptions. For instance, in the semantic web view ontologies model circumscribed interests, and do not make fundamental claims about reality. If some form of information integration is needed, alignment, extension and mapping are the actual targets, anything more being a plus. Unfortunately, this latter approach is not sufficient if the goal is to integrate the views of humans and robots for day-to-day cohabitation, possibly enhancing collaborations and social relationships. To be reliable, the ontology has to model how these agents understand reality.1 In the paper the term ‘understanding’ has a broad sense as it depends on the agent type. It covers notions like ‘building a model’, ‘attributing a behavior’ as well as ‘giving meaning’ to something.
Clearly, the environment is the same, and so is the material world. Or not? An embodied agent learns how the material world is by establishing relationships with the outside world: it learns and explores the environment via its body, cognitive capabilities, sensors. Robots lack cognition as humans know it, and their body is not only equipped with different sensors, it is not even biological. This makes robots’ experience of space, time and matter much different from that of humans. If the understanding of reality (whatever that means) depends on sensing and information processing, as we argue, different types of agent very likely develop different understandings of reality. How much different? That depends on the type of robot we are talking about and is, generally speaking, a complex question since it is unclear how to set up a possible comparison. Indeed, the initial claim that humans and robots are part of the same and only reality now sounds less reassuring.
What we are suggesting is that distinct agents (biological, artificial, cyborg) naturally develop distinct ontologies about what reality is, and that it is unlikely that different agents may agree on a unifying view, or even have the capability to understand each other’s systems. We should take seriously the possibility that humans and robots act according to views of reality that are incompatible, and perhaps largely incommunicable. We posit this as a problem for the future of our species and societies, and aim to indicate directions where ontological investigation is needed.
The rest of the paper is organized as follows: Section 2 introduces the distinction between situation and scenario, and characterizes the expression ‘self-taught robot’; Section 3 focuses on the change of perspective brought into applied ontology by the coexistence of highly heterogeneous robots; Section 4 shows the need to discuss interaction in a broader setting; and the concluding section, Section 5, points to a variety of related problems addressing briefly the world of non-human animals.
Due to the ongoing development of robotics, humans need to learn how to cope with robots. This is having an impact on human social behavior, and in particular on conventions [11], that at the moment seem discussed in terms of heteromation [7]. The norms that determine the organization of the human community and of human everyday interactions will adapt to emerging forms of robotics. However, AI methodologies like deep learning and hybrid symbolic-subsymbolic systems, make possible the creation of robots which achieve autonomy independently of human intervention, and this pushes us to imagine a variety of hypothetical situations. In this paper we are interested in situations that arise in a state of coexistence between humans and robots, the latter understood as embodied artificial agents. In particular, we look at situations in which humans and robots maintain their substantial independence and face the need to share space and resources to achieve their goals (survival, satisfaction of desires, ensuring safety conditions, adapting the environment to special needs).
Let us call
Let us call This should not be surprising since artificial agents have already the capability to learn how to cooperate via coordination in simple unsupervised settings [6].
Let us further constrain our focus by concentrating on self-robots that develop, among themselves, similar understanding of the environment, similar behavior and comparable goals provided their hardware is similar and, with respect to this, they live in comparable environments. Let us assume also that these self-robots are fairly coherent in matching behavior, scenarios, goals and actions. In different words, let us restrict our attention on self-robots that manifest regularities in dealing with the environment. This does not go as far as implying that, given a situation and the type of self-robot, they are predictable. The assumptions aim to ensure that it should be possible to develop some kind of shared set of rules and expectations among humans and self-robots, a set that can be the core for a system of norms [11] with which to regulate cohabiting and, perhaps, sustenance and cooperation among humans and self-robots.
Compared to humans, which have very similar bodies and capabilities, robots can show very heterogeneous characteristics: they can comprise different sensors and actuators, some even tailored to specific information or actions; have central or distributed processing units, process information locally, at the central level or in dedicated subsystems; reason with different computability and memory resources, and apply deduction rules, default rules, optimisation evaluations, probabilistic assignments, learned associations and so on, perhaps mixing these in a variety of ways. Each type of self-robot may understand a situation in a unique way but, due of the previous regularity assumption, it is still possible to ask what kind of understanding of reality such robot type may develop.
This question is not new. It rephrases the classical interoperability problem for which applied ontology has been introduced, that is, (
To cohabit with self-robots and perhaps interact with them for information exchange, collaboration, or simply to avoid ending up in dangerous situations, one needs to build a model of how robots understand reality and what needs they have, and possibly vice versa. The issue is not about merging or aligning ontologies in abstraction, for which there are different techniques in ontology engineering. The issue is whether today’s state-of-the-art in ontology engineering can make sense of ontological systems that are not human-based (leaving aside the related problem of how to elicit such ontologies).
To understand a self-robot which is physically and conceptually different from humans, that collects and classifies things in the environment according to viewpoints humans do not use or even consider, do humans need broader top-level ontologies (TLO) and a larger spectrum of agent-level ontologies (ALO)? While a TLO can be understood as a foundational ontology in the usual sense, a ALO is here seen as an agent’s specialization of a TLO. Thus, if the TLO is primarily influenced by the robot’s capabilities (e.g., to sense, reason and act), the ALO is the result of the ontological needs that the agent experience interacting with the outside world.
What would be an ontology that is not already covered by today’s TLO? A simple one, compatible with a robot that senses the environment at regular intervals, would claim that no event exists. There are only scenarios in which knowledge of objects and their properties, like relative position, is regularly override by the next sensing activity of the robot. This robot can develop a notion of causality, e.g., out of regularities in experienced sequences of scenarios. A notion of continuous change might be unaccessible to it (this depends on the relationship between sensing frequency and stimuli processing speed, compare the case of movie frames and vision processing in humans).
In another case the robot’s ontology may concentrate on material properties (density, thermal energy storage, chemical stability, etc.) disregarding features usually relevant in human TLOs like shape or even relations like connection. For instance, a robot that has sensors to detect compounds of fluorides, chlorides, nitrates and so on, could develop an ontology that centers around these distinctions and about substance concentrations in the environment.
These two cases are not really challenging. The first ontology, which to the best of our knowledge has not been formalized as a TLO in applied ontology, presents a reasonable view which is not seriously considered in our culture but has been philosophically considered and is at the core of some engineering devices. The second TLO is more likely a fragment of several existing TLOs. For instance, the DOLCE ontology [4] could be extended to include such an approach as an extension module of the Amount of Matter category.
More challenging for today’s TLOs are ontologies that rely on contextual information. Assume a self-robot uses its sensors to check if the environment is as desired. Whenever it discovers a difference between sensor data and expected data, it uses its actuators to change what is reachable in the direction where the problematic data are collected. Assume also that by doing this, it affects what it desires and thus the data it expects to perceive next. The challenge is that the robot is centred on a contextual classification of what it detects. From a standard perspective, the agent’s ALO may be seen as based on a few classes like a class of data and a class of areas, enriched with qualifications like perceived data vs goals, satisfactory vs unsatisfactory areas. However, the class of areas is not ontological in the standard sense as what an area is and how much it extends depends on the position and internal state of the robot, how the robot checks the environment (which sensors it decides to use), and how much the perceived data differ from the expected ones. One may develop a suitable TLO for such a robot but it would look quite different from existing TLOs. For instance, it would not make distinctions that today are generally adopted like among spatial region, physical object and individual quality. In short, this robot may behave rationally but its view of reality might not be compatible with TLOs today considered in applied ontology.
These examples of self-robot ontologies are quite simple and only aim to show unusual cases. The extent of human TLOs is further challenged when investigating robots based on neural network approaches: these may understand reality by detecting regularities in flows of information that overcome human capacities.
It might turn out that applied ontology, as we know it today, is fitted to develop TLOs suitable for self-robots, perhaps leaving out only some extreme cases, but most likely this is not so. We need to investigate methodologies that enable understanding of different ALOs, to integrate them within a larger class of TLOs, and to develop interfaces for information exchanges across these TLOs. The results of this line of research help to model the interaction between heterogenous agents like humans and robots, the focus of this paper, but also between different robot types and, to stretch it even further, between humans and aliens.
Interaction
This paper started with a practical motivation, expressed via hypothetical situations and scenarios, aiming to drive attention to the new ontological issues we are likely to face. The motivation, to foster possible interactions among human and artificial agents, points to another ontological problem: how should we understand the ontological notion of interaction?
The social systems that humans have experienced so far are either systems that evolved with humans, like cities, or are largely controlled by humans, like farms and hunting scenarios (with or without the support of non-human animals). These social systems are thus human-centred, and humans have been by far the most powerful agent in them. In the hypothetical world we are considering, humans may not have this special position. And this has important consequences.
The question that arises is not whether interaction in these hypothetical cases is possible, but in which sense it is. After all, one may doubt that humans and self-robots can make sense of each other’s behavior and expectations. Here the very use of the term ‘interaction’ may be challenged as it implicitly suggests some kind of purposefulness (at least from one of the interacting entities) combined with some form of reciprocity. Indeed, this is the standard understanding in robotics. In this sense, it brings to mind intentional agents. Even though agents capable of making decision and having goals may fit this view, it seems that the assumption of purposefulness and reciprocity is too restrictive.
I like to start from the notion of interaction as used in physics and engineering, essentially the way entities influence or affect each other’s behavior. Roughly, an interaction described at the level of physical laws, say the interaction between a book and the table where it lays, states that the behavior of one entity is influenced by the presence and behavior of the other. This physical interaction can be the starting point for an ontological analysis. Of course, not all interactions apply to entities controlled by physical laws only. A cognitive agent detecting an object in its environment, e.g., perceiving a car on its path, moves its attention to it (manifesting an interaction at the cognitive level) and may change position (interaction at the planning, functional and acting levels) to avoid contact. Other types of interaction occur at the social and cultural levels when joining a queue at an office or singing ‘Happy birthday’ at a party. In the human-computer interaction (HCI) domain there has been attempts to develop a technical language for modeling interaction [10] followed by efforts to ontologically generalize this view in other domains, e.g., [2], but the issue is clearly more general and of wider application.
The problem here is that the physical notion of interaction does not generalize well. Indeed, we have not developed a suitable framework for interaction beyond that of the physical laws. Without such framework we cannot establish how and to which extent an object may influence the behavior of another object. Yet, to make ontology suitable to evaluate interaction across humans and self-robots, a robust ontology of interaction must be developed.
Problems ahead of us
In the novel “The Book of Days”, Robert Chambers tells about a sow and her piglets charged and tried for the murder of a small child in 1457 [5]. Indicting a pig for a crime seems ridiculous today since, we believe, animals lack awareness of their actions and of the outcomes (interestingly, the book reports that “The sow was found guilty and condemned to death; but the pigs were acquitted on account of their youth, the bad example of their mother, and the absence of direct proof as to their having been concerned in the eating of the child.”3
From the point of view of modern animal studies as well as of robotics, the conclusion vary depending on the animal species, the robot system, and the theory of consciousness one takes [8,17]. The problem of founding and running a social and fair system that can comprise humans and self-robots remains widely open: we lack the basic principles about how to understand and organize such systems, and to measure their ethical status.
The previous observation triggers a series of topics, from issues rooted in cognitive and educational sciences to socio-technical organization and individual responsibility. Leaving these apart, we can imagine a classification of conditions for which social systems cannot even develop or survive beyond contingent or fortuitous circumstances. Among the foreseeable cases we can also imagine social systems that would very likely arise as an evolution of our existing systems. There is a need to study these conditions and how they relate to each other. On the ontological side, we should develop dedicated methodologies to understand the parameters to evaluate reliable interactions across different agent types, and even sustainable full fledged social systems.
Finally, most of the scenarios that we may want to elaborate about humans and self-robots are not completely new in our world. There are plenty of interactions between humans and other non-human animals which already give the gist of the ontological problems we are called to study. Nonetheless, there are two substantial differences between non-human animals and self-robots that push the problem at a different level. First, in modeling interactions humans have always taken the anthropological viewpoint. We have attributed to animals a worldview which apes our own. This move is not possible now. Second, human experience is limited to interactions with biological agents. For sure we can start from what we know about our mixed societies where human and non-human animals cohabit (with human consent or not) but we have to move beyond this if we want to be ready to develop a general social system or, at least, to reliably interact with self-robots and other beings.