Many highly sophisticated artificial intelligence

Many highly sophisticated artificial intelligence (AI) systems have already been created which by far surpass humans at the one particular task they are built for. However, those AI systems can only be used and tested in the context of their program or within simple virtual worlds. They are far too complex to operate in the real world and they are not equipped with the necessary sensors to communicate with or perceive the environment around them (Dominey and Warneken, 2009).
In contrast, robots are embodied systems that can interact with their environment with the help of sensors or the ability to move. Robots have long been established in various areas like industry, the military, or the medical field. However, most robots were also just able to perform the task they were created for, often to take over work that was to heavy or dangerous for human workers (Hockstein, Gourin, Faust & Terris, 2007). The first industrial robot was developed in 1959 for processing metal on an assembly line by carrying out the commands it was programmed with. (Hägele, Nilsson & Pires, 2008). This is an example of an approach called direct programming. The human engineer has to plan every step to get from the problem to the solution and program those steps into the robot. The robot is only able to follow the commands and carry out the tasks it was pre-programmed to do in the environment it was built for (Meeden & Blank, 2006). Another, more flexible approach is supervised learning. The human sets up different scenarios for the robot to learn how to act in or respond to different situations. The robot can generalize from the learnt behaviors and apply this knowledge to unknown situations that it did not practice before. However, this approach is limited to a small number of situations and the learnt actions can only be applied to similar tasks. Hence, it is not useful for inferring how to behave in a broad range of fundamentally different situations (Meeden & Blank, 2006).
To exhibit human-level social behaviour such as helping, the robot has to act on the same level as the human and needs to be able to think from the standpoint of another agent (Vernon, Thill & Ziemke, 2016). A more recent approach to achieving this is developmental robotics. One fundamental difference from the two previous approaches is that the engineer only equips the robot with very basic structures and then the robot learns by exploring and discovering its surroundings (Meeden & Blank, 2006; Faghihi & Moustafa, 2017). Cognitive robots could better navigate ambivalent and ill-structured situations and hypothesize about actions to be taken and their possible outcomes. Humans are cognitive agents, so to ensure comfortable interaction also robots need to have cognitive skills (Vernon et al., 2016). Developmental robotics suggests that it is often preferable to only equip the robot with the most elementary skills so that it can learn further behavior instead of programming a rigid robot with all behaviors it needs. In the first case, the robot is able to adapt to changing environments such as those when interacting with humans and can learn from and imitate human behavior to improve its social skills. This level of adaption is hardly possible in the second case (Pointeau & Dominey, 2017; Law, Shaw, Earland, Sheldon & Lee, 2014).
This newer approach to creating cognitive robots is based on the model of intelligent, living creatures, which also start off at a more basic and simple stage before over time developing their full potential (Meeden & Blank, 2006; Law et al., 2014; Asada et al., 2009). Developmental psychology describes and researches how human beings develop their intelligence and learn to display complex social behaviors, among others. An important theoretical assumption of developmental psychology is that complex social behavior does not happen at once, but in stages (Law et. al, 2014; Piaget, 1952). The advantage is that at each stage only a relatively simple behavior has to develop, and then each subsequent stage can build on what was previously learnt. The shift from one stage to the subsequent one is not abrupt, but smoothly happens over time (Piaget, 1952). Ultimately, each step in development adds up to the complex social behavior that can be seen in humans (Law et. al, 2014; Asada et al., 2009). Since a long-term goal in robotics and also artificial intelligence has been to create a general intelligence as well as human-like social behavior and cognition, the aim of this paper is to analyze concepts of developmental psychology that aid in creating more complex cognitive functions and improved social behaviour in robots and how this in turn can benefit developmental psychology.
Implementation of developmental psychological concepts in cognitive robots
The approach of developmental robotics is closely related to that of developmental psychology: we are born, or in the case of robots, created, with some basic abilities like perception. New, advanced knowledge and abilities are then generated by interacting with and imitating those in our environment. However, to implement those fundamental abilities in a robot, there has to be some rudimental answer to the question which abilities we are born with and which ones we learn later. Multiple studies agree that necessary requirements for more complex development are perception and movement to be able to engage with the environment, as well as language to communicate and understand commands, and an autobiographical memory (ABM) to encode these different experiences (Spelke & Kinzler, 2007, Lallée et al., 2011, Pointeau & Dominey, 2017).
An important aspect in the process of reaching full human social behavior is the development of a self or sense of self. The self is, for instance, a basic requirement for evolving the ability to cooperate with other social agents (Pointeau & Dominey, 2017). According to Neisser (1988), the self is not a whole but is split into different entities, which all follow a different developmental process. The ecological self is the individual in interaction with its direct environment and already emerges early in infancy. The interpersonal self emerges at around the same time and is the self that is communicating and socially interacting with another individual. The extended self consists of our experiences and memories as well as habits and hopes or anticipations for the future. The conceptual self starts developing in the second year of life and is the representation that an individual has of him- or herself. This can include for example social roles or character traits, and not everything that is part of the conceptual self is necessarily true. The private self emerges when children discover that they have private thoughts, feelings and experiences, that no other person has access to (Neisser, 1988).
Autobiographical memory and the experience stored in it are crucial for developing a self or sense of self (Fivush, 2011; Schacter, Welker & Wang, 2016). Autobiographical memory structures new experiences and thus aids the development of a self. It is suggested that robots implemented with an autobiographical memory can to some degree develop Neisser’s first four levels of self. Achieving these different types of selves is an important goal or even necessity for developmental robotics (Pointeau ; Dominey, 2017). In humans, experience encoded in long-term memory can be useful in learning as well as inferring what to do in unknown situations (Kolb, 2014). Therefore, the ABM cannot only be a storage for experiences. The encoded memories need to be meaningful, which means that it has to be possible to make changes to the ABM or its structure and add new memories or knowledge. Only then can the ABM aid in the development of a self. However, the robot not only has to be able to encode experiences, but also to generate conclusions from regular patterns in those memories with the help of reasoning. These generalizations can then help to choose the right actions in unknown conditions (Pointeau ; Dominey, 2017).
The ecological self is the first of Neisser’s selves that develops in humans and it is fundamental for the development of the other parts of the self. Before a sense for social interactions and the knowledge of past and future experience can emerge, one has to learn about one’s own body and its relation to the immediate environment. To achieve that, Pointeau and Dominey (2017) let the iCubLyon01 robot that they used in their study learn a body schema as well as a model of its actions. In this study, the learning attempts of the robot were mainly instructed by a human teacher. However, Schillaci, Hafner and Lara (2016a) have shown that a robot can also learn to form a body image by individually exploring its environment.
Similar to the way a child develops, the robot is based on algorithms that learn about actions by comparing the states before and after the action with the help of basic sensations like touch or movement (Pointeau, Petit ; Dominey, 2014). By storing those memories, the robot can find regular patterns and thus learn to anticipate possible consequences of actions it performs. With that knowledge, the robot can create a simulated model, called forward model (Schillaci, Ritter, Hafner ; Lara, 2016b). When carrying out an action, the actual outcome is compared to the expected model based on experience, and when they differ this information can be used to adjust the action (Pointeau, Petit ; Dominey, 2013).
To get closer to realizing the ecological self, the robot should also have a representation of its own body. Research in neuroscience suggests an embodied model built around convergence-divergence zones (CDZs). Signals from separate motor and sensory areas in the brain or robotic system merge in integrative areas that make up the CDZs, which is what happens during the process of learning. To recall the whole sensory experience, only one modality has to be activated. This in turn activates the CDZ, which will diverge activation to the other representations involved (Meyer ; Damasio, 2009; Jouen et al., 2015). This type of architecture was also implemented in the iCub robot in the form of self-organizing multi-modal convergence maps (MMCM). This should help the robot to form a connection between his different senses, by merging their different inputs of the same object or stimulus in the CDZ. The input from the various senses was repeatedly paired in the CDZ while the robot performed different actions, thus enabling it to form a sensory-motor image of its body (Pointeau ; Dominey, 2017).
With the help of the forward model that is realized in the ABM and the body schema, the robot can make sense of how its actions affect the environment around it as well as its own body, hence leading closer towards implementing an ecological self in the robot (Pointeau ; Dominey, 2017; Schillaci et al., 2016b). To have a functioning ABM, it is required that the agent has at least a simple form of first person perspective (1PP). This basic 1PP seems to be present in the iCub, since it has formed an embodied representation of its own body within the environment. This is closely related to the concept of the ecological self, which refers to a person interacting with its environment (Pointeau ; Dominey, 2017).
The interpersonal self is responsible for human social interaction, which includes spoken language, body language, and facial expressions, among others. Also, cooperation is a feature of the interpersonal self. To be able to cooperate, oneself and the individual one wants to act with need to agree on a shared plan, which encompasses the steps both individuals need to take to jointly reach the shared goal (Tomasello, Carpenter, Call, Behne ; Moll, 2005). Hence, cooperating with others is closely interwoven with the interpersonal self. The forward model that was used to predict the iCub’s own actions could also be used to create a model of another’s action and compare the real outcome to the outcome predicted by the model. In this way, the iCub could also tell whether someone was lying about the action the individual announced to perform (Pointeau et al., 2013).
A similar study (Copete, Nagai ; Asada, 2016) found that learning the pairing of one’s own sensory and motor signals when carrying out an action could be useful for predicting the observed actions of others. Research by Hafner and Kaplan (2008) has found that when acting together with others, agents can learn to add the actions of others to the models of their own actions or to include those other agents in their own body representation. This shows that in cooperative action, oneself and the other are closely linked. The results from the previously mentioned studies helped the robot to take part in basic interaction consisting of a single action.
A shared plan contains multiple actions, which all have to be carried out in a given order by two or more cooperating agents to be able to reach a common goal. The actions in the shared plan include variable arguments, for example objects that have to be moved. Once the plan is learned, it can also be performed with new objects (Pointeau & Dominey, 2017). The iCub learned the shared plan with the help of instructions and demonstrations by the human. While encoding the shared plan in the ABM, the system performed argument matching, which means that objects or agents are recognized as variables at their different occurrences in the plan (Lallée et al., 2013). As a result, they can be exchanged for other objects or agents when later carrying out the shared plan, enabling the robot to generalize from one stored plan to multiple different situations. One trial of instruction and demonstration was enough for the robot to learn the shared plan (Pointeau & Dominey, 2017).
For a robot to cooperate successfully with a naïve human, it has to disclose its shared plan to the human with a combination of language and gestures. The interpersonal self is important both for sharing one’s own desires as well as for understanding actions and social expressions of others. The ABM is necessary to store and order relevant experience for the interpersonal self. In the ABM the instructions of the different actions are combined to form the shared plan, which could then be carried out in cooperation with the human (Pointeau ; Dominey, 2017).
According to Neisser (1988), the development of the conceptual self is highly dependent on experience and represents the view that we have of ourselves and our personality. We obtain our concepts both through observation and instruction, for example from our parents, or more generally, from our culture. The development of these higher-level self-concepts depends on the accumulation and consolidation of the experiences made in developing the ecological and the interpersonal self, both in humans and in robots. The knowledge of the robot concerning spatial locations and action prediction relating to the development of the ecological self can be expanded to general rules. Those aid in dealing with unknown situations, which is related to and required for the development of the conceptual self (Pointeau ; Dominey, 2017; Petit, Pointeau ; Dominey, 2016).
An important factor in developing a conceptual self is to use the concept of oneself to take the perspective of someone else, which is a fundamental requirement for achieving a theory of mind. To test this, the well-known Sally-Anne task can be used, which was created to test false beliefs and perspective taking in children (Baron-Cohen, Leslie ; Frith, 1985). The child sees the two dolls, Sally and Anne. Sally puts her toy in a basket, and then walks away. Anne takes the toy and puts it in a box. The child is then asked where Sally will look for the toy when she comes back. The answer depends on whether the child is able to take Sally’s perspective or not (Baron-Cohen et al., 1985).
Humans are capable of creating mental models or simulations of ourselves and our actions, which allows us to move around in time and space by remembering the past or envisioning the future. This property may be due to the ability of forward modeling, which is also given by the ABM (Pointeau & Dominey, 2017). From the experience encoded in the ABM, the iCub can form general rules that enable it to predict its own as well as the human’s actions. This model is created quicker than the actual action is carried out, so the outcome of the action can be compared to the simulated outcome. The simulated model can be used to check one’s own actions or predict those of another, or additionally to arrange prospective actions or take the perspective of someone else (Pointeau & Dominey, 2017).
The iCub is equipped with an Objects Properties Collector (OPC), which represents the human working memory and contains all knowledge the robot has about an ongoing situation. The iCub also possesses a second mental OPC (MOPC) which is used to review and examine experiences in order to encode them in the ABM. This architecture was useful in performing the Sally-Anne experiment. When an object was put into the first position by a human agent acting as Sally, the MOPC was updated to take the same state as the OPC. Sally left and the iCub paused the MOPC to represent Sally’s world view. A second agent representing Anne moved the object to a different location. The iCub had the accurate view that the object was now in a new location, but that Sally thought that it was still in its first location, as stored in the paused MOPC (Pointeau ; Dominey, 2017). These findings show that the iCub robot is moving close to developing a conceptual self as described by Neisser (1988).
The extended self is similar to the concept of the narrative self as described by Bruner (2009), which implies that narrative organizes the communication with other individuals across different time periods. The shared plans that the iCub learned in cooperation with the human agent are closely related to the concept of the extended self. This is because these plans include interaction with another individual and can be used again and again over time. The ABM of the iCub has been recording and storing its memories of various actions, encounters, and interactions for four years. These experiences could be added as parts of a shared plan, which could then be used repeatedly. This in combination worked towards developing a narrative self (Pointeau et al., 2014). Due to the system architecture of the robot, it can retrieve its memories and at the same time communicate those to a human interaction partner (Pointeau ; Dominey, 2017).
Another important factor in creating a narrative self is narrative enrichment, which means associating experiences with each other that are not obviously connected. Instead of representing each event separately in the ABM, they are connected among each other and linked with their featured agents with the help of a situation model (Dominey, Mealier, Pointeau, Mirliaz ; Finlayson, 2017). In other words, narrative can connect separate experiences into a story and help to illustrate the overall context.
Interacting with other social agents is an important part of developing human social abilities. For these interactions to function, the social agent first needs to develop a self. As suggested by the findings described above, a crucial aspect in the development of the self is memory, especially ABM. These studies show how the ABM can aid in developing the different parts of the self as described by Neisser (1988). In favor of developmental psychology, this supports existing theories on the significance of autobiographical memory in the development of social cognition. For developmental robotics, these findings identify a way of creating improved social cognition in robots. Furthermore, there is a reciprocal relationship between the development of the self and social interaction, with one depending on the other to be able to form (Pointeau ; Dominey, 2017).
Social cognition develops over time and is fundamental for socially interacting with others, so robots which are supposed to live in a human world need to be capable of social cognition. To be capable of social cognition, agents have to process incoming sensory information about the intentions and actions of the agents they interact with (Vernon et al., 2016). To ensure functioning social cognition and comfortable interaction between agents, it is important that the agent can identify motion and interpret gestures and body language (Vernon et al., 2016). This ability is already present in newborns and aids in adapting to the interaction partner and in communicating nonverbally (Simion, Regolin ; Bulf, 2008). One way to achieve this is to implement a World Model in the system, which stores the current state of the world. This world model is connected to and can be changed by the robot’s vision, its motor abilities and some form of a language processing system (Dominey & Warneken, 2011). The cognitive state of an agent is connected with its bodily state, so the body reveals much about the agent’s thoughts and feelings and can also influence the behaviour or reactions of another agent.
Social cognition is made up of helping, which is one-directional, and collaboration, which is bidirectional. Collaboration is underpinned by the capability of helping behavior and humans need several years to fully develop these capacities (Vernon et al., 2016). Developing motor skills as well as determining others’ focus of attention helps in learning to grasp the intentions of others (Gredebäck & Kochukhova, 2010). Starting from the second year of life, children develop the capacity for instrumental helping, they help without being asked to and also when no reward is offered (Warneken & Tomasello, 2009; Vernon et al., 2016).
Collaboration, which is reciprocal, implies that the two agents have to combine their efforts and be sensitive to the actions of the other agent to appropriately adjust their own actions. The two agents want to reach a shared goal with the help of shared intentions, joined attention and with the use of joint action, all contained in a shared plan (Tomasello & Carpenter, 2007; Vernon et al., 2016). Those plans can be saved in an Intentional Plan Repertory (IntRep) in the form of sequential actions that contain involved agents, objects and goal of the action (Dominey & Warneken, 2011). Both agents need to know not only the steps they themselves have to take to reach the goal, but also the actions of the other agent, so they can support the other when needed or adjust their actions at the right time. For recognizing a shared plan, the actions of the other agent are compared to the sequences stored in the IntRep. If the ongoing action equals the start of a stored plan, the robot knows which plan is being carried out. As a result, the robot can predict successive actions or take over if the other agent fails to carry out its part of the plan (Dominey & Warneken, 2011). To successfully reach the shared goal, the agents need to be attentive to the actions of each other and have to be willing to immediately step in to help the other with completing their steps if needed (Vernon et al., 2016). This helping occurs instinctively as a result of the robot’s goal to complete the shared plan. If the other agent fails, the robot knows what to do by accessing the plan and can ask to take over the action of the other agent (Dominey ; Warneken, 2011).
In order to accomplish a shared goal, each agent has to be able to predict the outcome of its own actions as well as the actions and their outcomes of the other agent (Bratman, 1992; Butterfill, 2012). The agents need also be aware how they influence the behaviour of the other agent and that some steps can only be solved by joined effort of the two agents (Sebanz ; Knoblich, 2009). The agents need to adjust their actions to continually move towards completion of the shared goal and support the efforts of the other agent if necessary (Vernon et al., 2016; Pacherie, 2013; Butterfill, 2012). Movements and actions differ in the intentions they are associated with – low-level and high-level intentions, respectively. Low-level intentions explain which goal or what the agent wants to reach, and high-level intention explains for what reason or why the agent acts that way (Pacherie, 2013).
Reading intentions is related to the concept of theory of mind. Theory of mind is uniquely human and refers to the ability of seeing from another’s perspective and recognizing and comprehending the other agent’s thoughts, emotions or desires (Meltzoff, 1995). Learning through observation and imitation is fundamental for a person to develop a theory of mind, which is in turn crucial for evolving cognitive functions (Meltzoff ; Decety, 2003; Meltzoff ; Moore, 1997). Already toddlers younger than two years of age can infer a goal that an adult has in mind by imitating and pursuing the action the adult initiated but failed to complete, thus deducing the adult’s intention (Meltzoff & Decety, 2003). Cognitive systems deduce others’ intentions in the same way as inferring the outcomes of their own actions. These systems use simulations where the input is a motor command or observed action of another agent and the output are the possible outcomes of performing those movements or observed actions (Iacoboni, 2009; Ondobaka ; Bekkering, 2012). Consequently, with the help of these inputs and outputs, the ultimate goal of the other agent can be deduced from these actions.
Another important factor in children’s development, that gets them to explore the world around them, and as a result learn about it, is intrinsic motivation (Law et al., 2014). Some behavior is driven by external motivation, such as rewards. However, this can only account for a small fraction of behavior and motivation. With internal motivation, the child acts without any external reinforcement or structure. Without external structure or guidance, the child does not know which one out of a number of possible actions to chose if they are all equally plausible (Law et al., 2014).
One approach to implement intrinsic motivation in robots is to use novelty to simulate intrinsic motivation and guide exploration and learning. This is achieved by invoking interest for all unknown events. The expression ‘unknown events’ is meant extensively and includes internal as well as external events, such as stimuli from muscles or senses, different types of interactions, or old events in a new order (Law et al., 2014). All experiences or stimuli are characterized as new if they have not been encountered previously or if the robot failed to anticipate them. This relates back to the autobiographical memory described in the beginning. Events that have already been encountered need to be stored and given the ability to be manipulated to recognize and correctly deal with unknown experiences (Law et al., 2014).
If a stimulus is encountered multiple times, the robot habituates to it and loses interest in the stimulus. In the beginning of the robot’s development, simple actions or objects are new and stimulating but as the robot develops, it adapts to those and then more complex objects or interactions move to the focus of interest (Law et al., 2014). Regarding technical realization, the different events or stimuli are first given a high weight to represent their high excitation. The weights are controlled by a function which initially increases sensitivity to the stimulus but then decreases the weight every time the stimulus is encountered again, thereby decreasing its excitation (Law et al., 2014). A majority of studies in the developmental robotics context focus primarily on one stage of development, such as the development of the self described above. However, Law et al. (2014) suggest that it might be more beneficial to implement development over the long term. They argue that each stage builds up on its preceding one and that each stage can thus affect behavior emerging from the successive stage (Law et al., 2014).
Implementing developmental theories in cognitive robots not only benefits the advance of robot cognition but can also benefit developmental psychology. To be able to develop an algorithm and program a developmental model into a robot, it needs to be very precise and detailed (Dominey ; Warneken, 2011). This can encourage researchers to create well-structured theories that are suitable to be tested and can produce relevant findings (D’Mello & Franklin, 2011). Furthermore, it is often time-consuming or complicated to study developmental theories. Cognitive robots can provide an accessible way for testing those theories and for examining how development is related to and influenced by the environment (Dominey & Warneken, 2011). If they can be successfully implemented in a robot, that in itself is an indicator whether the model is reliable and functioning or not. Additionally, the behavior of the robot can be observed and can be used to supplement empirical research findings (Kelley & Cassenti, 2011, D’Mello ; Franklin, 2011).
Discussion
The aim of this paper was to analyze concepts of developmental psychology that aid in creating more complex cognitive functions and improved social behaviour in robots and how this in turn can benefit developmental psychology. Developmental psychology is a broad field, therefore taking into consideration all concepts of this field would be far beyond the scope of this thesis.
The first part examined the development of a self as described by Neisser (1988) and how this can be implemented in a robot with the help of an autobiographical memory structure. The autobiographical memory is used for storing and manipulating experiences. The robot can then use these experiences in learning first about its own body when developing the ecological self. These schemas are later extended to encompass other agents and more complex interactions when evolving the other levels of self (Pointeau ; Dominey, 2017). Moreover, the paper dealt with how with the help of shared plans and cooperation, the robot can develop the ability to infer others’ intentions as well as show helping behaviour similar to the one observed in children (Dominey & Warneken, 2011). Additionally, it was examined how intrinsic motivation and novelty can guide the robot’s exploration of the environment and encourage it to learning new and more complex actions that build on the ones that were learned in a previous stage (Law et al., 2014). Lastly, benefits of developmental robotics for developmental psychology were discussed, for instance that developmental cognitive robots can be useful for testing and improving developmental psychological models (D’Mello & Franklin, 2011).
These findings show that, with the help of developmental psychological theories, robots can learn complex social behavior similar to the one expressed in young children. This is an advantage over direct programming and can produce robots that can flexibly adapt to the dynamic human environment. This world is becoming increasingly automated, with robots becoming more important in various different areas. They can, for instance, take over basic jobs in social care, where there is always a shortage of human workers. This could improve the working conditions for the human workers by taking away some of the pressure they are subject to. Also, the robots can ensure that the patients are always well cared for, even if only few human workers are available.
Due to the benefits for developmental psychology and also psychology in general, cognitive social robots can help to further the understanding of human cognition, thus deepening the understanding of our own brains and minds. As a result, better theories and experiments might be designed, which could ultimately lead to better treatments of various kinds of mental or cognitive problems. Furthermore, developmental robots are only equipped with basic structures, therefore requiring less engineering and programming effort from the part of the human. Additionally, because the robot is not made for a specific task, it can be taught various tasks depending on what it is needed for. It can learn different skills or actions depending on the environment it is free to explore or depending on the different types of social interactions it encounters. This can save production costs or efforts by only having to produce one type of robot which can then be shaped to learn what is needed.