An architecture for ethical robots inspired by the simulation theory of cognition

The expanding ability of robots to take unsupervised decisions renders it imperative that mechanisms are in place to guarantee the safety of their behaviour. Moreover, intelligent autonomous robots should be more than safe; arguably they should also be explicitly ethical. In this paper, we put forward a method for implementing ethical behaviour in robots inspired by the simulation theory of cognition. In contrast to existing frameworks for robot ethics, our approach does not rely on the veriﬁcation of logic statements. Rather, it utilises internal simulations which allow the robot to simulate actions and predict their consequences. Therefore, our method is a form of robotic imagery. To demonstrate the proposed architecture, we implement a version of this architecture on a humanoid NAO robot so that it behaves according to Asimov’s laws of robotics. In a series of four experiments, using a second NAO robot as a proxy for the human, we demonstrate that the Ethical Layer enables the robot to prevent the human from coming to harm in simple test scenarios. (cid:1) 2017 The Authors. Published by Elsevier B.V. Thisisan open accessarticleunder the CCBYlicense(http://creativecommons.org/licenses/by/4.0/).


Introduction
Robots are becoming ever more autonomous.Semiautonomous flying robots are commercially available, and driver-less cars are undergoing real-world tests (Waldrop, 2015).This trend towards robots with increased autonomy is expected to continue (Anderson & Anderson, 2007).An expanding ability to take unsupervised decisions renders it imperative that mechanisms are in place to guarantee the safety of behaviour executed by the robot.The fact that many robots are designed to interact with humans further heightens the importance of equipping robots with mechanisms guaranteeing safety (Royakkers & van Est, 2015;Winfield, 2012).For example, the state-of-the-art in robots for care, companionship, and collaborative manufacturing is rapidly advancing (Goeldner, Herstatt, & Tietze, 2015;Lin, Abney, & Bekey, 2011).At the other end of the spectrum of robot-human interaction, the development of fully autonomous robots for military applications is progressing rapidly (e.g., Arkin, Ulam, & Wagner, 2012;Lin et al., 2011;Sharkey, 2008;Xin & Bin, 2013).
Robot safety is essential but not sufficient.Smart autonomous robots should be more than safe; they should also be explicitly ethical -able to both choose and justify (Anderson & Anderson, 2007;Moor, 2006) actions that prevent harm.As the cognitive, perceptual and motor capabilities of robots expand, they will be expected to have an improved capacity for moral judgment.As summarised by Picard and Picard (1997), the greater the freedom of a machine, the more it will need moral standards.
The necessity of robots equipped with ethical capacities is recognised both in academia (e.g., Arkin et al., 2012;Deng, 2015;Gips, 2005;Moor, 2006;Picard & Picard, 1997;Wallach & Allen, 2008) and wider society, with influential figures such as Bill Gates, Elon Musk and Stephen Hawking speaking out on the dangers of increasing autonomy in artificial agents.Nevertheless, only a few studies have implemented robot ethics.To the best of our knowledge, the efforts of Anderson and Anderson (2010) and our previous work (Winfield, Blum, & Liu, 2014) are the only instances of robots equipped with (limited) moral principles.So far, most work has been either theoretical (e.g., Mackworth, 2011;Wallach & Allen, 2008) or simulation based (e.g., Arkin et al., 2012).Irrespective of whether the work was done in real robots or not, existing architectures for ethical robots are based on logic frameworks (Arkin et al., 2012;Bringsjord, Arkoudas, & Bello, 2006;Govindarajulu & Bringsjord, 2015).This approach uses artificial reasoning processes to verify whether the robotic behaviour satisfies a set of predetermined ethical constraints.This approach to ethical robots is reminiscent of Good Old-Fashioned AI (GOFAI) in the sense that it relies heavily on abstract symbolic reasoning (Mackworth, 2011).

The simulation theory of cognition
Pinker (1997) argued extensively that the human mind has not evolved to be an abstract symbol manipulator.Since then, advances in cognitive science have confirmed that the computations underlying human cognition are very different from rule-based manipulation of abstract symbols (Barsalou, 2010).This view has emerged in many domains of human cognition, including perception, reasoning and problem-solving (see Barsalou, 1999;Dijkstra & Post, 2015;Hegarty, 2004;Wilson, 2002 for more examples).Moreover, representing, learning and combining concepts leads to some problems in purely symbolic systems (Ga ¨rdenfors, 2004;Lieto, Chella, & Frixione, 2016).Therefore, it seems the mind uses representations that are richer than the abstract symbols allowed for in models of intelligence that presume abstract symbols.
The theory of mind that allows for the richest representations is the simulation theory of cognition (Hesslow, 2002;Wilson, 2002).It hypothesises that thinking utilises the same cognitive (and neural) processes as interaction with the external environment.When thinking, actions are covert and are assumed to generate, via associative brain mechanisms, the sensory inputs that elicit further actions (Hesslow, 2012).In this view, thinking requires building a grounded model of the environment -which is not composed of abstract symbols.Rather, it is assumed to re-instantiate and recombine experiences using the brain's systems of perception, action, and emotion.The mental model covertly simulates actions and their associated perceptual effects (see Hegarty, 2004;Hesslow, 2002;Hesslow, 2012;Wilson, 2002 for reviews).
In this paper, we put forward a method for implementing ethical behaviour in robots inspired by the simulation theory of cognition.In contrast to existing frameworks for robot ethics, our approach does not rely on the verification of logic statements.Rather, it utilises internal simulations which allow the robot to simulate actions and predict their consequences.Therefore, our method is a form of robotic imagery.Many other areas of robotics have exploited robotic imagery.In their review of robotic imagery, Marques and Holland (2009) coined the term functional imagination to denote the mechanism whereby robots covertly simulate actions and their consequences to steer their future behaviour.Here we adopt their term.Hence, this paper aims at advancing functional imagination as a method for ethical robots.
We aim at implementing consequentialist ethics, which is implicit in the very common conception of morality, shared by many cultures and traditions (Haines, 2015).Hence, developing an architecture suited for this class of ethics is a reasonable starting point.Moreover, the primary advantage of a functional imagination is the ability to test the outcome of potential actions (Hesslow, 2002;Hesslow, 2012) without committing to them (Marques & Holland, 2009;Ziemke, Jirenhed, & Hesslow, 2005).Therefore, functional imagination is a framework suitable for supporting consequentialist ethics.

Architecture
Over the years, keeping track with shifts in paradigms (Murphy, 2000), many architectures for robot controllers have been proposed (see Kortenkamp & Simmons (2008, 2005, 2000) for reviews).However, given the hierarchical organisation of behaviour (Botvinick, 2008), most robotic control architectures can be remapped onto a threelayered model (Kortenkamp & Simmons, 2008).In this model, each control level is characterised by differences in the degree of abstraction and time scale at which it operates.At the top level, the controller generates long-term goals (e.g.'Deliver the package to room 221').Next, goals are translated into a set of tasks that should be executed (e.g.'Follow corridor', 'Open door', etc.).Finally, the tasks are translated into (sensori) motor actions that can be executed by the robot (e.g.'Raise arm' and 'Turn wrist joint').Obviously, this general characterization ignores many particulars of individual control architectures.
Assuming that the robot is controlled by a three-layered controller (Fig. 1a), we agree with Arkin (Arkin, 2008;Arkin et al., 2012) that ethical behaviour should be governed by adding a fourth specialised control layer.This Ethical Layer (Fig. 1b) should act as a governor evaluating behaviour proposed by each of the three other layers before the robot executes it.In principle, the functionality of the Ethical Layer could be distributed across and integrated with the layers present in existing control architectures.Indeed, in humans, ethical decision making is most likely supported by the same computational machinery as decision making in other domains (Young & Dungan, 2012).Nevertheless, from an engineering point of view, guaranteeing the ethical behaviour of the robot through a separate