Infant-like Social Interactions between a Robot and a Human Caregiver

From birth, human infants are immersed in a social environment that allows them to learn by leveraging the skills and capabilities of their caregivers. A critical pre-cursor to this type of social learning is the ability to maintain interaction levels that are neither overwhelming nor under-stimulating. In this paper, we present a mechanism for an autonomous robot to regulate the intensity of its social interactions with a human. Similar to the feedback from infant to caregiver, the robot uses expressive displays to modulate the interaction intensity. This mechanism is integrated within a general framework that combines perception, attention, drives, emotions, behavior selection, and motor acts. We present a specific implementation of this architecture that enables the robot to react appropriately to both social stimuli (faces) and non-social stimuli (moving toys) while maintaining a suitable interaction intensity. We present results from both face-to-face interactions and interactions mediated through a toy. Infant-like Social Interactions between a Robot and a Human Caregiver p.3 Infant-like Social Interactions between a Robot and a Human Caregiver Social robotics has generally concentrated on the behavior of groups of robots performing behaviors such as flocking, foraging or dispersion (Balch & Arkin, 1994; Mataric, 1995) or on paired robot-robot interactions (Billard & Dautenhahn, 1997). Our work focuses not on robot-robot interactions, but rather on the construction of robots that engage in meaningful social exchanges with humans. By doing so, it is possible to have a socially sophisticated human assist the robot in acquiring more complex communication skills and in learning the meaning these acts have for others. The interactions with the caregiver can bootstrap the robot s capabilities. By leveraging the skills and abilities of a benevolent caregiver, it is possible to alleviate many of the normal difficulties of robot learning, such as sparse reinforcement, unconstrained task complexity, and unstructured environments. Our approach is inspired by the way infants learn to communicate with adults. An infant s emotions and drives play an important role in generating meaningful interactions with the caregiver (Bullowa, 1979). These interactions constitute learning episodes for new behaviors. In particular, the infant is strongly biased to learn communication skills that result in having the caregiver satisfy the infant s drives (Halliday, 1975). The infant s emotional responses provide important cues which the caregiver uses to assess how to satiate the infant s drives, and how to carefully regulate the complexity of the interaction. The former is critical for the infant to learn how its actions influence the caregiver, and the later is critical for establishing and maintaining a suitable learning environment for the infant. A critical pre-cursor to this type of social learning is the ability to maintain interaction levels that are neither overwhelming nor under-stimulating. In this paper, we Infant-like Social Interactions between a Robot and a Human Caregiver p.4 present a mechanism for an autonomous robot to regulate the intensity of its social interactions with a human. This mechanism is the first stage of a long-term endeavor to enable social learning between the robot and a human caregiver and is integrated within a general framework that combines perception, attention, drives, emotions, behavior arbitration, and motor acts (Breazeal, 1998). We concentrate on the design specification of the perceptual and motivational systems because of the critical role they serve in this dynamic process for infants. Other work in progress focuses on the construction of shared attention systems that allow the infant and the caregiver to ground learning in perceptual episodes (Scassellati, 1996, 1998c). The specifics of the learning algorithms have yet to be addressed. We do not claim that this system models infant development. However, the design is heavily inspired by the role motivations and facial expressions play in social interaction between infants and adults. Regulating interaction intensity is a critical skill for this kind of social learning because it helps the caregiver tune her actions so that they are appropriate for the infant. For our purposes, the context for learning involves social exchanges where the robot learns how to manipulate the caregiver into satisfying its internal drives. Ultimately, the communication skills targeted for robot learning are those exhibited by infants, such as turn taking, shared attention, and pre-linguistic vocalizations exhibiting shared meaning with the caregiver. This paper is organized as follows: first we discuss the numerous roles motivations play in natural systems—particularly as they apply to behavior selection, regulating the intensity of social interactions, and learning in a social context. Next we describe a robot called Kismet that has been designed and built to provide emotional feedback to the caregiver through facial expressions. We then present a framework for the design of the behavior engine, which integrates perception, motivation (drives and emotions), attention, behavior, and motor skills (expressive or task based). Particular detail is provided for the design of the perceptual and motivational systems. After we illustrate these ideas with a specific Infant-like Social Interactions between a Robot and a Human Caregiver p.5 implementation on a physical robot, we present the results of some early experiments in which a human engages the robot in face-to-face social exchanges. 1. THE ROLE OF MOTIVATIONS IN SOCIAL INTERACTION Motivations, which encompass drives, emotions, and pain, play several important roles for both arbitrating and learning behavior. We are interested in how they influence behavior selection, regulate social interactions, and promote learning in a social context. Behavior Selection Much of the work in motivation theory in ethology is intended to explain how animals engage in appropriate behaviors at the appropriate time to promote survival (Lorenz, 1973; Tinbergen, 1951). Internal drives influence which behavior the animal pursues. Furthermore, the same sensory stimulus may result in very different behavior depending on the intensity of the drives. For example, a dog will respond differently to a bone when it is hungry than when it is fleeing from danger. It is also well accepted that animals learn things that facilitate the achievement of biologically significant goals. Motivations provide an impetus for this learning. In particular, the motivational system provides a reinforcement signal that guides what the animal learns and in what context. When an animal has a strong drive that it is trying to satisfy, it is primed to learn behaviors that directly act to satiate that drive. For this reason, it is much easier to train a hungry animal than a satiated one with a food reward (Lorenz, 1973). For a robot, an important function of the motivation system is to regulate behavior selection so that the observable behavior appears coherent, appropriately persistent, and relevant given the internal state of the robot and the external state of the environment. The responsibility for this function falls largely under the drive system of the robot. Other work in autonomous agent research has used drives in a similar manner (Arkin, 1988; Maes, 1992; Infant-like Social Interactions between a Robot and a Human Caregiver p.6 McFarland & Bosser 1993; Steels 1995). Drives are also necessary for establishing the context for learning as well as providing a reinforcing signal. Blumberg (1996) used motivations (called internal variables) in this way to implement operant conditioning so that a human user could teach an animated dog new tricks. Regulating Interaction An infant s motivations are vital to regulating social interactions with the caregiver (Kaye, 1979). Soon after birth, an infant is able to display a wide variety of facial expressions (Trevarthen, 1979). As such, the infant responds to events in the world with expressive cues that can be read, interpreted, and acted upon. The caregiver interprets them as indicators of the infant s internal state (how he or she feels and why), and acts to promote the infant s well being (Chappell & Sander, 1979; Tronick, Als, and Adamson, 1979). For example, when the infant appears content the caregiver tends to maintain the current level of interaction, but when the infant appears disinterested the caregiver intensifies or changes the interaction to try to re-engage the infant. In this manner, the infant can regulate the intensity of interaction by displaying appropriate emotive cues. The caregiver instinctively reads the infant s expressive signals and acts to maintain a level of interaction suitable for him. An important function for a robot s motivational system is not only to establish appropriate interactions with the caregiver, but also to regulate the intensity so that the robot is neither overwhelmed nor under-stimulated. When designed properly, the intensity of the robot s expressions provide appropriate cues for the caregiver to increase the intensity of the interaction, decrease the intensity, or maintain it at the current level. By doing so, both parties can modify their own behavior and the behavior of the other to maintain the intensity of interaction that the robot requires. Infant-like Social Interactions between a Robot and a Human Caregiver p.7 Learning in a Social Context The use of emotional expressions and gestures facilitates and biases learning during social exchanges. Caregivers take an active role in shaping and guiding how and what infants learn by means of scaffolding. As the word implies, the caregiver provides a supportive framework for the infant by manipulating the infant s interactions with the environment to foster novel abilities. Commonly, scaffolding involves reducing distractions, marking the task s critical attributes, reducing the number of degrees of freedom in the target task, providing ongoing reinforcement through expressive displays of face and voice, or enabling the infant to experience the outcome of a sequence of activities before the infant is cognitively or physically able to attain it for himself or herself (Wood, Bruner, and Ross, 1976). The emotive cues that the adult receives during social exchanges serve as feedback so that the adult can adjust the nature and intensity of the structured learning episode to maintain a suitable learning environment in which the infant is neither bored nor overwhelmed. In addition, during early interactions with the caregiver, an infant s motivations and emotional displays are critical in establishing the foundational context for learning episodes (Halliday, 1975). An infant displays a wide assortment of emotive cues such as coos, smiles, waves, and kicks during early face-to-face exchanges. During the first month, the infant s basic needs, emotions, and emotive expressions are among the few things the adult thinks they share in common. Consequently, the caregiver imparts a consistent meaning to the infant s expressive gestures and expressions, interpreting them as meaningful responses and as indications of the infant s internal state. Curiously, experiments by Kaye (1979) argue that the caregiver actually supplies most if not all of the meaning to the exchange when the infant is very young. The infant does not know the significance that expressive acts have for the adult, nor how to use them to evoke specific responses. However, because the adult assumes the infant shares the same Infant-like Social Interactions between a Robot and a Human Caregiver p.8 meanings for emotive acts, this consistency allows the infant to discover what sorts of activities will get specific responses. Routine sequences of a predictable nature can be built up, which serve as the basis of learning episodes (Newson, 1979). Furthermore, they provide a context of mutual expectations. For example, early cries of an infant elicit various caregiving responses, depending upon how the adult initially interprets these cries and how the infant responds. The infant and the adult converge over time on specific meanings for different kinds of cries. The infant comes gradually to differentiate his or her cries (i.e., cries of distress, cries for attention, cries of pain, cries of fear) in order to elicit different responses from the caregiver. The adult reinforces the shared meaning of the cries by responding in consistent ways to these variants. Evidence of this differentiation is provided by the development of unique communication protocols that differ from those of other adult-infant pairs (Bullowa, 1979). Combining these ideas, a robot can be biased to learn how its emotive acts influence the caregiver in order to satisfy its own drives. Toward this end, we endow the robot with a motivational system that works to maintain its drives within homeostatic bounds and a set of emotive expressions analogous to the types of emotive expressions that human infants display. These capabilities allow the caregiver to observe the robot s emotive expressions and interpret them as reflections of the robot s internal drives. The human can then act appropriately. This interaction establishes the routine necessary for the robot to learn (eventually) how its emotive acts influence the behavior of the caregiver, and how these acts ultimately serve to satiate the robot s own drives. This section has argued that motivations should play a significant role in determining the robot s behavior, how it interacts with the caregiver, and what it can learn during social exchanges. With these long term challenges in mind, an important pre-requisite function for the robot s motivational system is not only to establish appropriate interactions with the human, but also to regulate the interaction intensity so that the robot can learn without being overwhelmed or under-stimulated. When designed properly, the interaction among the Infant-like Social Interactions between a Robot and a Human Caregiver p.9 robot s drives, emotions, and expressions provide appropriate cues for the caregiver so that he or she knows whether to change the activity itself or to modify its intensity. By doing so, both parties can modify both their own behavior and the behavior of the other in order to maintain an interaction from which the robot can learn from and use to satisfy its drives.