Is imitation learning the route to humanoid robots?

This review investigates two recent developments in artificial intelligence and neural computation: learning from imitation and the development of humanoid robots. It is postulated that the study of imitation learning offers a promising route to gain new insights into mechanisms of perceptual motor control that could ultimately lead to the creation of autonomous humanoid robots. Imitation learning focuses on three important issues: efficient motor learning, the connection between action and perception, and modular motor control in the form of movement primitives. It is reviewed here how research on representations of, and functional connections between, action and perception have contributed to our understanding of motor acts of other beings. The recent discovery that some areas in the primate brain are active during both movement perception and execution has provided a hypothetical neural basis of imitation. Computational approaches to imitation learning are also described, initially from the perspective of traditional AI and robotics, but also from the perspective of neural network models and statistical-learning research. Parallels and differences between biological and computational approaches to imitation are highlighted and an overview of current projects that actually employ imitation learning for humanoid robots is given.     

Predictive coding as a model of cognition

Previous work has shown that predictive coding can provide a detailed explanation of a very wide range of low-level perceptual processes. It is also widely believed that predictive coding can account for high-level, cognitive, abilities. This article provides support for this view by showing that predictive coding can simulate phenomena such as categorisation, the influence of abstract knowledge on perception, recall and reasoning about conceptual knowledge, context-dependent behavioural control, and naive physics. The particular implementation of predictive coding used here (PC/BC-DIM) has previously been used to simulate low-level perceptual behaviour and the neural mechanisms that underlie them. This algorithm thus provides a single framework for modelling both perceptual and cognitive brain function.     

Inhibition of return in perception and action

Inhibition of return (IOR) refers to the delay in responses to previously cued locations. Whether IOR influences perceptual and/or motor processes has been controversial. To determine IOR effects on perception and action, this study examined IOR in spatially directed hand reaching (Experiment 1) and spatial localization of targets with a mouse cursor (i.e., an indirect visuomotor mapping/perceptual task; Experiment 2). The reaction times showed delayed responses for targets appearing within the whole cued hemifield for both tasks. However, hypometric spatial biases were consistently found only with directed reaching. Spatial biases in the mouse localization task were indirectly influenced by IOR and distinct from those in the reaching task. The dissociation in spatial characteristics for directed reaching vs. perception suggests that the effects of IOR are task dependent, but may be more directly linked to the dorsal motor system.     

The role of sensorimotor experience in the development of mimicry in infancy

During social interactions we often have an automatic and unconscious tendency to copy or ‘mimic’ others’ actions. The dominant view on the neural basis of mimicry appeals to an automatic coupling between perception and action. It has been suggested that this coupling is formed through associative learning during correlated sensorimotor experience. Although studies with adult participants have provided support for this hypothesis, little is known about the role of sensorimotor experience in supporting the development of perceptual‐motor couplings, and consequently mimicry behaviour, in infancy. Here we investigated whether the extent to which an observed action elicits mimicry depends on the opportunity an infant has had to develop perceptual‐motor couplings for this action through correlated sensorimotor experience. We found that mothers’ tendency to imitate their 4‐month‐olds’ facial expressions during a parent‐child interaction session was related to infants’ facial mimicry as measured by electromyography. Maternal facial imitation was not related to infants’ mimicry of hand actions, and instead we found preliminary evidence that infants’ tendency to look at their own hands may be related to their tendency to mimic hand actions. These results are consistent with the idea that mimicry is supported by perceptual‐motor couplings that are formed through correlated sensorimotor experience obtained by observing one's own actions and imitative social partners.     

A review of abnormalities in the perception of visual illusions in schizophrenia

Specific abnormalities of vision in schizophrenia have been observed to affect high-level and some low-level integration mechanisms, suggesting that people with schizophrenia may experience anomalies across different stages in the visual system affecting either early or late processing or both. Here, we review the research into visual illusion perception in schizophrenia and the issues which previous research has faced. One general finding that emerged from the literature is that those with schizophrenia are mostly immune to the effects of high-level illusory displays, but this effect is not consistent across all low-level illusions. The present review suggests that this resistance is due to the weakening of top–down perceptual mechanisms and may be relevant to the understanding of symptoms of visual distortion rather than hallucinations as previously thought.     

Action-induced effects on perception depend neither on element-level nor on set-level similarity between stimulus and response sets

As was shown by Wykowska, Schubö, and Hommel (Journal of Experimental Psychology, Human Perception and Performance, 35, 1755–1769, 2009), action control can affect rather early perceptual processes in visual search: Although size pop-outs are detected faster when having prepared for a manual grasping action, luminance pop-outs benefit from preparing for a pointing action. In the present study, we demonstrate that this effect of action–target congruency does not rely on, or vary with, set-level similarity or element-level similarity between perception and action—two factors that play crucial roles in standard stimulus–response interactions and in models accounting for these interactions. This result suggests that action control biases perceptual processes in specific ways that go beyond standard stimulus–response compatibility effects and supports the idea that action–target congruency taps into a fundamental characteristic of human action control.     

The composite face illusion: A whole window into our understanding of holistic face perception

Two identical top halves of a face are perceived as being different when their bottom halves belong to different faces, showing that the parts of a face cannot be perceived independently from the whole face. When this visual illusion is inserted in a matching task, observers make more mistakes and/or are slower at matching identical top face halves aligned with different bottom halves than when the bottom halves are spatially offset: The composite face effect. This composite face paradigm has been used in more than 60 studies that have provided information about the specificity and nature of perceptual integration between facial parts (“holistic face perception”), the impairment of this process in acquired prosopagnosia, its developmental course, temporal dynamics, and neural basis. Following a review of the main contributions made with the paradigm, I explain its rationale and strengths, and discuss its methodological parameters, making a number of proposals for its optimal use and refinement in order to improve our understanding of holistic face perception. Finally, I explain how this standard composite face paradigm is fundamentally different than the application to facial parts of a congruency/interference paradigm that has a long tradition in experimental psychology since Stroop (1935), and which was originally developed to measure attentional and response interference between different representations rather than perceptual integration. Moreover, a version of this congruency/interference paradigm used extensively over the past years with composite faces lacks a baseline measure and has decisional, attentional, and stimulus confounds, making the findings of these studies impossible to interpret in terms of holistic perception. I conclude by encouraging researchers in this field to concentrate fully on the standard composite face paradigm, gaze contingency, and other behavioural measures that can help us take one of the most important challenges of visual perception research: Understanding the neural mechanisms of holistic face perception.     

Grasp cueing shows obligatory attention to action goals

To understand the grounding of cognitive mechanisms in perception and action, we used a simple detection task to determine how long it takes to predict an action goal from the perception of grasp postures and whether this prediction is under strategic control. Healthy observers detected visual probes over small or large objects after seeing either a precision grip or a power grip posture. Although the posture was uninformative it induced attention shifts to the grasp-congruent object within 350 ms. When the posture predicted target appearance over the grasp-incongruent object, observers' initial strategic allocation of attention was overruled by the congruency between grasp and object. These results might help to characterize the human mirror neuron system and reveal how joint attention tunes early perceptual processes toward action prediction.     

具身概念表征的研究及理论述评

具身认知强调认知在本质上是具身的, 身体在认知的实现中发挥着关键作用。传统的符号加工理论认为, 概念表征独立于主体的知觉运动系统并以抽象符号的形式储存于语言记忆中。概念表征的具身理论则认为, 概念表征与知觉运动系统具有共同的神经基础, 概念在本质上是主体经验客体时知觉与运动体验的神经记录, 而概念加工的基本形式则是身体经验的模拟与还原。关于该理论的实证研究主要集中于概念加工引发的知觉动作变化、身体动作对概念加工的影响、抽象概念加工的具身特征等领域。今后的研究应关注符号加工理论与具身理论的整合等。     

Independence and integration of perception and action: An introduction

This special issue examines the relationships between cognitive systems for perception and action. “Action” is meant in a very broad sense, to include processes of selecting, planning, and executing overt responses. Recent interest in the psychology and neuropsychology of action has led to a variety of approaches describing visuomotor systems and the relationship between perception and action. There is at least one basic constraint on this relationship that everyone can agree on: Perceptual systems have evolved to guide action. So although it is at least conceivable that perceptual processing could go on largely independent of concurrent action, a system that planned and executed actions without perceptual guidance would be worse than useless — it would be a complete disaster in anything but the most predictable environment.     

Depth perception by the active observer

The connection between perception and action has classically been studied in one direction only: the effect of perception on subsequent action. Although our actions can modify our perceptions externally, by modifying the world or our view of it, it has recently become clear that even without this external feedback the preparation and execution of a variety of motor actions can have an effect on three-dimensional perceptual processes. Here, we review the ways in which an observer's motor actions--locomotion, head and eye movements, and object manipulation--affect his or her perception and representation of three-dimensional objects and space. Allowing observers to act can drastically change the way they perceive the third dimension, as well as how scientists view depth perception.     

Gestalt principles in the control of motor action

We argue that 4 fundamental gestalt phenomena in perception apply to the control of motor action. First, a motor gestalt, like a perceptual gestalt, is holistic in the sense that it is processed as a single unit. This notion is consistent with reaction time results indicating that all gestures for a brief unit of action must be programmed prior to initiation of any part of the movement. Additional reaction time results related to initiation of longer responses are consistent with processing in terms of a sequence of indivisible motor gestalts. Some actions (e.g., many involving coordination of the hands) can be carried out effectively only if represented as a unitary gestalt. Second, a perceptual gestalt is independent of specific sensory receptors, as evidenced by perceptual constancy. In a similar manner a motor gestalt can be represented independently of specific muscular effectors, thereby allowing motor constancy. Third, just as a perceptual pattern (e.g., a Necker cube) is exclusively structured into only 1 of its possible configurations at any moment in time, processing prior to action is limited to 1 motor gestalt. Fourth, grouping in apparent motion leads to stream segregation in visual and auditory perception; this segregation is present in motor action and is dependent on the temporal rate. We discuss congruence of gestalt phenomena across perception and motor action (a) in relation to a unitary perceptual-motor code, (b) with respect to differences in the role of awareness, and (c) in conjunction with separate neural pathways for conscious perception and motor control.     

Contrasting conceptions of perception and action

An examination of contrasting conceptions of perceptual processes, control of action and the nature of the linkage between perception and action leads to identification of a family of options. Two of these options, labeled pure realism and pure constructivism, are widely recognized. The argument is offered that the other options, which are hybrid networks of propositions, cannot be ruled out on a priori grounds. Nevertheless, the hybrids are usually ignored. It is proposed that investigators in the field of perception and action should embark on programmatic research that evaluates these contrasting conceptions on common empirical terrain.     

Observing or producing a motor action improves later perception of biological motion: Evidence for a gender effect

Two experiments are presented addressing the issue of whether observing (visual priming) or producing (motor priming) a running activity during a very short period (30 s) facilitates the perception of the direction of a point-light runner embedded in a dense dynamical mask. Experiment 1 showed that perceptual judgements improved and response time increased in the visual priming compared to the neutral priming condition (video of a moving car) in which judgements were at random. Because this effect was observed for male participants only, we performed a second experiment with the aim of evaluating the role of gender congruency in the visual priming condition. Results confirmed the facilitation effect and demonstrated that this effect was strictly dependent on the gender congruency between the perceiver and the priming information. Moreover, we found that actually producing a motor activity similar to the one presented in the video sequence improved to the same extent participants’ judgement of the direction of the point-light runner, without any gender effect. As a whole, these findings argue in favour of common representation for the perception and the production of human movement and showed that the perception of biological motion can be improved by prior motor activity either performed or observed. However, the gender-dependent effect of visual priming suggested that motor repertoire differed in males and females.     

An enhanced experimental procedure to rationalize on the impairment of perception of action capabilities

It is well documented that changes in the physiological states of the perceiver–actor influence the perception of action capabilities. However, because experimental procedures of most studies involved a limitless availability for stimuli visual encoding and perceptual strategies, it remains difficult to adopt a single position among the large range of alternative interpretations for impaired perception. A reaching-to-grasp paradigm under breathing restriction was adapted from Graydon et al. (Cogn Emot 26:1301–1305, 2012) to standardize the time for encoding of stimuli information and narrowed the involvement of perceptual strategies. In the present study, we propose a highly controlled environment where the discrete information is presented during 300 ms, congruently with neurophysiological studies focused on visuomotor transformation. An underestimation of the perception of action capabilities is found under breath restriction, suggesting that 300 ms for stimuli encoding is sufficient to induce altered visuomotor brain transformations when limiting the involvement of perceptual strategies. This result suggests that such behavior could refer to an impaired brain potentiation of the perceptual occurrence, providing strong hypotheses on the brain dynamics of sensorimotor integration that underlie impaired perception of action capabilities in stressful situations.     

Impact of action planning on spatial perception: Attention matters

Previous research suggested that perception of spatial location is biased towards spatial goals of planned hand movements. In the present study I show that an analogous perceptual distortion can be observed if attention is paid to a spatial location in the absence of planning a hand movement. Participants judged the position of a target during preparation of a mouse movement, the end point of which could deviate from the target by a varying degree in Exp. 1. Judgments of target position were systematically affected by movement characteristics consistent with perceptual assimilation between the target and the planned movement goal. This effect was neither due to an impact of motor execution on judgments (Exp. 2) nor due to characteristics of the movement cues or of certain target positions (Exp. 3, Exp. 5A). When the task included deployment of attention to spatial positions (former movement goals) in preparation for a secondary perceptual task, an effect emerged that was comparable with the bias associated with movement planning (Exp. 4, Exp. 5B). These results indicate that visual distortions accompanying manipulations of variables related to action could be mediated by attentional mechanisms.     

Mechanisms of selection for the control of hand action

Most attention research has viewed selection as essentially a perceptual problem, with attentional mechanisms required to protect the senses from overload. Although this might indeed be one of several functions that attention serves, the need for selection also arises when one considers the requirement of actions rather than perception. This review examines recent attempts to determine the role played by selective mechanisms in the control of action. Recent studies looking at reach-to-grasp responses to target objects in the presence of distracting objects within a three-dimensional space are discussed. The manner in which motor aspects of the reach-to-grasp response might be influenced by distractors is also highlighted, rather than merely addressing the perceptual consequences of distractors. The studies reviewed here emphasize several factors highlighting the importance of studying selective processes within three-dimensional environments from which attention and action have evolved.     

Perceptual resonance: action-induced modulation of perception

A direct relationship between perception and action implies bi-directionality, and predicts not only effects of perception on action but also effects of action on perception. Modern theories of social cognition have intensively examined the relation from perception to action and propose that mirroring the observed actions of others underlies action understanding. Here, we suggest that this view is incomplete, as it neglects the perspective of the actor. We will review empirical evidence showing the effects of self-generated action on perceptual judgments. We propose that producing action might prime perception in a way that observers are selectively sensitive to related or similar actions of conspecifics. Therefore, perceptual resonance, not motor resonance, might be decisive for grounding sympathy and empathy and, thus, successful social interactions.     

Perceiving while acting: action affects perception

In two experiments we studied how motor responses affect stimulus encoding when stimuli and responses are functionally unrelated and merely overlap in time. Such R-S effects across S-R assignments have been reported by Schubö, Aschersleben, and Prinz (2001), who found that stimulus encoding was affected by concurrent response execution in the sense of a contrast (i.e., emphasizing differences). The present study aimed at elucidating the mechanisms underlying this effect. Experiment 1 studied the time course of the R-S effect. Contrast was only obtained for short intertrial intervals (ITIs). With long ITIs contrast turned into assimilation (i.e., emphasizing similarities). Experiment 2 excluded an interpretation of the assimilation effect in terms of motor repetition. Our findings support the notion of a shared representational domain for perception and action control, and suggest that contrast between stimulus and response codes emerges when two S-R assignments compete with each other in perception. When perceptual competition is over, assimilation emerges in memory.