ABiALS 2010/11: Spatial Representations and Dynamic Interactions
Poster Abstracts
Acoustic Grounding of Spatial Frames of Reference ? Influence of response actions on space region concepts
Marcella de Castro Campos, Bettina Bläsing, Thomas Herrmann, Constanze Vorwerg & Thomas Schack
Spatial language can give us insights into the nature of spatial cognition in general. Spatial region concepts such as front, back, right and left have characteristics that reflect our typical interaction with space, mostly regarded in the visual domain (Franklin, Henkel & Zangas, 1995). Humans are primarily visually oriented; in absence of visual information, spatial orientation and representation is mainly guided by audition (Blauert, 1997). In this project we are interested in relations between spatial region concepts, their grounding in the acoustic domain and their dependency on response actions. In our first experiment, we have analysed verbal and action responses from 24 subjects indicating sound source locations (SSL). While standing in the middle of a ring of 16 loudspeakers, each subject had to respond to acoustic signals (3 finger snaps) displayed by single speakers in three conditions: by (A) verbal indication of the SSL using predefined labels (front, back, left, right), (B) verbal indication of the SSL plus turning the head and upper body towards the stimulus, and (C) verbal indication of the SSL plus turning towards the stimulus plus pointing with the hand and outstretched arm. Verbal labels given in each trial were recorded and coded directly, and a 6 camera VICON motion capture system monitored head orientation and pointing movement in space via retro-reflective markers fixed on the subject?s head, arms and upper body. The data are currently being analysed. We expect that spatial concept regions will differ between the three conditions and from those obtained in studies using visual stimuli (Franklin et al., 1995). The findings will contribute to research regarding spatial representations and auditory processing, and will be of interest for human orientation in real-world scenarios, in human computer interaction, ambient computing and environments for patients with impaired hearing or sight (e.g., sports facilities).
Blauert, J. (1997). Spatial Hearing: the psychophysics of human sound localization. Massachusetts, USA: The MIT Press.
Franklin, N., Henkel, L.A. & Zangas, T. (1995). Parsing surrounding space into regions. Memory & Cognition, 23 (4), 397-407.
Goal-directed Action Understanding via Coupled Internal Models: A Probabilistic Approach
Haris Dindo
Many activities performed in daily life involve some form of interaction and collaboration with other people. Consider for instance the tasks of jointly moving a table or building an IKEA furniture from scratch. These tasks require that distinct actors coordinate their actions at different levels of complexity, from low-level motor features to high-level \mental" states dealing with others' goals, intentions and beliefs. An imperative view in social cognition literature indicates that such a coordination is achieved by the process of action simulation in which one's own internal structures used to actively interact with the external world are employed for understanding of our conspecifics.
Computational approaches of this idea are mostly based on coupled forward-inverse internal models indicated as core ingredients for motor control and used for explaining complex social abilities in humans, ranging from action understanding to theory of mind. However, for each task there might be hundreds or thousands of internal models and it is unlikely and impractical that all models are maintained in parallel. Problems associated with the number of internal models to take into consideration during action simulation, and their inherent stochastic nature, hinder the development of efficient analytical solutions.
I will describe a generative Bayesian model for action simulation, in which inverse-forward model pairs are considered \hypotheses" of plausible action goals that are explored in parallel via an approximate inference mechanism. We adopt particle filters, a Monte Carlo technique for sequential simulation, where each particle represents a weighted hypothesis of an internal model activation in the action simulation task. The reenactment of inverse-forward model pairs is a form of action simulation, which supports both perceptual prediction and action understanding at the goal level. In three action perception experiments, we test how prior information initially biases the action simulation process, and how it gets successively refined when novel evidence is collected that confirms or disconfirms the predictions generated by the models. Our results show that motor simulation permits dynamically reallocating the particles depending on the prediction accuracy of the inverse-forward models, and leads to successful goal-directed action understanding in scenarios resembling those adopted in social cognitive neuroscience studies.
Cognitive and motor capacities of ARoS
Wolfram Erlhagen, Estela Bicho, Luis Louro, Eliana Costa e Silva
A long-term goal of our research group is to build robots that are able to interact with users in the same way as humans interact with each other in common tasks. Our approach to achieve this demanding goal is to take inspiration from experimental and theoretical findings in studies that investigate social interactions of humans and other species.
Here we present results of our ongoing research on efficient and fluent human-robot collaboration that is heavily inspired by recent experimental findings about the neuro-cognitive mechanisms supporting joint action in humans. The robot control architecture implements the joint coordination of actions and goals as a dynamic process that integrates contextual cues, shared task knowledge and the predicted outcome of the user's motor behavior. The architecture is formalized as a coupled system of dynamic neural fields representing a distributed network of local but connected neural populations with specific functionalities. We validate the approach in a task in which the humanoid robot ARoS and a human user jointly assemble toy objects. We show that the context-dependent mapping from action observation onto appropriate complementary actions allows the robot to cope with dynamically changing joint action situations. More specifically, the results illustrate crucial cognitive capacities for efficient and successful human-robot collaboration such as goal inference, error detection and anticipatory action selection.
To translate the abstract decisions about adequate complementary actions into overt motor behavior we apply a posture-based planning approach inspired by the knowledge model of Rosenbaum and colleagues. It allows us to generate human-like smooth and collision-free movements that may be easily interpreted by a human observer as goal-directed.
Structural Knowledge Transfer by Spatial Abstraction for Learning Agents
Lutz Frommberger
This contribution investigates the role of abstraction principles for knowledge reuse in agent control learning tasks. It is shown how aspectualization, a particular form of abstraction, can be utilized to achieve knowledge transfer in reinforcement learning. The use of so-called structure space aspectualizable knowledge representations explicates structural properties of the state space and enables "a posteriori structure space aspectualization" (APSST) as a method to extract generally sensible behavior from a learned policy. This new policy can be used for knowledge transfer to support learning new tasks in different environments. A case study demonstrates transfer of generally sensible navigation skills from simple simulation to a real-world robotic platform.
Prospective coding of external-spatial location in touch
Tobias Heed, Johanna Möller & Brigitte Röder
Movement planning and execution influence stimulus processing in many ways. For example, when planning a saccade, neurons with visual receptive fields at the location at which a target will be located after the saccade fire already before movement execution; the brain thus represents the future location of the visual stimulus. Similarly, stimulus processing is enhanced at the goal location of upcoming saccades and arm movements. Here we asked whether the perceived location of tactile stimuli is influenced by the coordinates of planned hand movements. When participants judge the order of two tactile stimuli, one to each hand, they make many errors when the hands are crossed, presumably because both skin-based and external coordinates are used to represent tactile location; in crossed postures, these reference frames signal conflicting spatial information. Participants made reaching movements with both hands, starting and ending in uncrossed and crossed postures. Two tactile stimuli were presented either before or during reaching, and participants indicated which of the two stimuli came first. Tactile performance was impaired when movements started in a crossed posture, indicating the known coordinate conflict between somatotopic and external coordinates for tactile localization. Importantly, performance was impaired also when movements ended in a crossed posture, even when stimulus presentation preceded hand movements. Thus, tactile stimulus location is recoded into external spatial coordinates not only according to the current, but also according to the planned body posture, indicating that the predicted posture expected to result from a planned movement is used to compute tactile spatial location. Such prospective coding in external coordinates may be useful to plan a motor response to the tactile location, in that this movement can be planned towards the future location of the touched limb already before the movement into the new posture is completed.
Can an age-related recalibration of perceptual and action-related representations of the perceptuomotor workspace account for differences in visuomotor adaptation between younger and older adults? Thoughts and preliminary evidence
Mathias Hegele & Herbert Heuer
Adaptation to a visuomotor transformation is known to be impaired at older adult age. More specifically, it has been shown for both, gain and rotation adaptation, that the age-related impairment primarily pertains to conscious strategic corrections and the explicit knowledge on which they are based, but not to the (implicit) acquisition of an internal model of the transformation. The age-related deficit in visuomotor adaptation has been associated with the decay of various cognitive functions in older age, such as decision-making or spatial working memory. Here, we argue that it might also be of importance to take into account the influence of represented metrics of the perceptuomotor workspace. Data from our lab provide preliminary support for this notion. In a study investigating age-related differences of gain adaptation, we observed an age-related dissociation of explicitly judged and actually executed amplitudes of hand movements. While both groups judged the distance of required hand movements similarly, older adults produced significantly shorter movement amplitudes than their younger counterparts. Moreover, in a recent study of rotation adaptation, we observed a residual age-related deficit in adaptive shifts despite comparable amounts of explicit knowledge due to prior cognitive training of the rotation. While younger adults were able to use this knowledge to fully compensate the rotation during goal-directed reaching movements, older adults still undercompensated the rotation in actual movement execution. We suggest that these results reflect a recalibration of motor and perceptual space in the elderly, which limits their application of explicit knowledge represented in perceptual space through strategic corrections of hand movements in motor space, thereby reducing overall visuomotor adaptation to a novel transformation. Results are discussed with reference to studies indicating that aging affects the representation of peripersonal sensorimotor space.
Robotic self-models
Matej Hoffmann
How is our body imprinted in our brain? Despite substantial efforts, the mysteries of body representations are far from uncovered. The most widely used notions?body image and body schema?are still waiting to be clearly defined. The mechanisms that underlie body representations are co-responsible for the admiring capabilities that humans or many mammals can display: combining information from multiple sensory modalities, controlling their complex bodies, adapting to growth, failures, or using tools. These features are also desirable in robots. We review the concept of body schema in robotics. First, we examine application-oriented research: how a robot can improve its capabilities by being able to automatically synthesize, extend, or adapt a model of its body. Second, we summarize the research area in which robots are used as tools to verify hypotheses on the mechanisms underlying biological body representations.
Computational models of Social Learning and Exploration
Manuel Lopes
Social learning provides a vast amount of information that can be used by learners. Computational models have been developed to understand how new tasks, new primitives and new goals can be acquired from demonstrations and interactions. Social learning is influenced by several factors: beliefs about the world's possible states and actions causing transitions between them; baseline preferences for certain actions; a variable tendency to infer and share goals in observed behaviour; and a variable tendency to act efficiently to reach rewarding states. In this work we address these issues by considering a computational model that can incorporate all these factors. We also explore how learning from demonstration, or social learning in general, can be seen as a form of active exploration of the environment. We show how a learner can efficiently request demonstrations from a teacher. These interactions can occur even when the feedback from the demonstrator is ambiguous or not known.
Predicting the Focus of Attention and Deficits in Situation Awareness with a Modular Hierarchical Bayesian Driver Model
Claus Möbus & Mark Eilers
Situation Awareness (SA) is defined as the perception of elements in the environment within a volume of time and space, the comprehension of their meaning, and the projection of their status in the near future. Lacking SA or having inadequate SA has been identified as one of the primary factors in accidents attributed to human error.
In this poster we present a probabilistic machine-learning-based approach for predicting the focus of attention and deficits of SA in real-time using a Bayesian driver model as a driving monitor. This Bayesian driving monitor generates expectations concerning the focus of attention and deficits in SA conditional on the actions of the driver.
Buying Time: Optimally Planning while Acting and Monitoring in a Dynamic Stochastic Environment
Dimitri Ognibene & Giovanni Pezzulo
We analyze the problem of how an agent can anticipate the proper allocation of its computational resources to prepare proper plans while being immersed in a dynamic stochastic environment. In our formulation the agent is the intermediary between two stochastic processes: the environment and the planning process.
Since in this work we are mainly interested in the interaction between these stochastic processes, we will assume that the environment is partitioned in a set of independent and qualitatively equivalent sub-environments which are connected each other in a stochastic manner. The agent will plan for the next sub-environments it may face.
In this condition, planning during the interaction with the environment allows the agent to focus only on the most relevant future sub-environments because the agent has updated information on the current state and the actual results of its actions. Both what sub-environment the agent will finally face and when are stochastic variables. An agent that allocates its resources optimally should predict the evolution of these variables and the costs of facing a situation with an approximated plan.
The planning processes for each sub-environment are themselves stochastic thus the agent has to predict their evolution and estimate how much and when it will gain from refining the plan for a specific sub-environment.
By taking into account the utility of the time needed for planning for future environments, the agent can capitalize being in a ``safe'' condition by preparing itself before to face a critical sub-environment. On the other side the agent can try to avoid sub-environments for which its planning process is going slower than expected. This, we hypothesize, is the base mechanism underlying the so called epistemic actions, actions executed in order to facilitate computations.
Motor resonance after learning to use a rattle: Electrophysiological evidence for action-effect binding in 8-month-old infants
Markus Paulus, Sabine Hunnius, Michiel van Elk & Harold Bekkering
Bidirectional action-effect associations play a fundamental role in intentional action control and the development of the mirror neuron system. However, little is known about the developmental origins of the acquisition of bidirectional action-effect associations. To investigate this, we trained 8-month-old infants for one week to use a novel rattle that produced a specific sound when shaken. During this week infants were also repeatedly presented with another sound, which was not related to an action. After this training phase, infants? EEG responses to these two sounds and an additional, unfamiliar sound were recorded. The results show that infants displayed a stronger mu-desynchronization above cortical motor sites (i.e. motor resonance) when listening to the action-related sound than when hearing other sounds. Our results provide therefore the first direct evidence that infants as young as 8 months of age are able to acquire bidirectional action-effect associations and parallel findings of audiovisual mirror neurons in the monkey brain.
Distributed model for sensorimotor control: anticipatory coordination and lateral competition
Jean-Charles Quinton
We here propose a two layer modular infrastructure and evaluate various algorithms for competition and spatiotemporal coordination in order to control artificial sensorimotor systems. This research is part of a broader project that aims at understanding the mechanisms and constraints necessary to the emergence of adaptive behaviors from acquired distributed representations. The proposed model takes inspiration from the cerebral cortex organization at a mesoscopic scale, but targets computationally efficient implementations.
The upper layer is composed of distributed spatiotemporal representations, making predictions about the dynamics of the agent and its environment based on the context and actions taken. These anticipations continuously monitor their adequacy with the current sensorimotor flow, which determines their level of activity and influence on the overall dynamics. They are thus task-independent and can be learned in an unsupervised way. They can however flexibly coordinate through a distance based back-propagation of activity between them, as to produce goal-oriented trajectories.
The resulting future-oriented activity is projected on the lower layer, that implements competition between bottom-up and top-down signals. As sensory and motor signals have been unified in the upper layer, this layer not only produces an interpretation of the sensory flow but also dynamically selects the most adequate actions. Although various algorithms have been explored, lateral inhibitory connection schemes are the most promising as they allow attentional properties to emerge and make the system highly robust to noise.
The topology introduced in both layers can be refined by introducing modular maps manipulating only a subset of sensorimotor dimensions, and considering a generalized propagation function between partially specified states. These conceptual extensions are coupled with software and hardware optimization techniques as to provide real-time interaction capabilities.
Learning of time delays with and without haptic demonstration
Katrin Rapp & Herbert Heuer
Movements often serve to produce expected effects in the environment, as for instance when pushing a button to open the door of a train. When the effect does not become visible immediately, the movement tends to be repeated. Despite this and other detrimental performance effects of delays, manual control with delayed effects can in principle gain from practice. Also, recent results hint at a favourable effect of haptic guidance on learning of motor timing. In the present experiment we compared practice with and without haptic guidance on learning a pure time delay in a step-tracking task. In addition to performance during practice, we compared the changes of performance in test conditions without haptic guidance and without visual feedback around the occurrence of a target step, both with and without the temporal delay. In line with typical findings, the immediate performance effect of robot guidance was clearly present. However, when guidance was no longer available, the beneficial effect of guidance disappeared. Adaptive shifts and after-effects showed no advantage of the haptic-guidance group. To the contrary, with the non-significant advantage of the no-guidance group increasing from 38 ms and 70 ms for intermediate test and post test, it might well be, that with extended practice a reliable disadvantage of the guidance group could develop. Altogether, our results question the hypothesis that haptic demonstration could support the learning of motor timing.
Motor Simulation: How an action-perception-overlap affects internal action prediction
Anne Springer, Simone Brandstädter & Wolfgang Prinz
Previous studies provided evidence for the claim that the prediction of occluded action involves an internal real-time simulation. We present two experiments that aimed to study how real-time simulation is affected by simultaneous action execution under conditions of full, partial or no overlap between observed and executed actions. This overlap was analyzed by comparing the body sides and the movement kinematics involved in the observed and the executed action. While performing actions, participants observed point-light actions that were interrupted by an occluder, followed by a test pose. The task was to judge whether the test pose depicted a continuation of the occluded action in the same depth angle. Using a paradigm proposed by Graf et al. we independently manipulated the duration of the occluder and the temporal advance of the test pose relative to occlusion onset (occluder time and pose time, respectively). This paradigm allows to assessing real-time simulation, based on prediction performance across different occluder time/pose time combinations (i.e., an increase in task performance with decreasing time distance between occluder time and pose time is taken to reflect real-time simulation). The point-light actor could be perceived from the front or the back, as indicated by task instructions. In Experiment 1 (front view instructions), evidence of action simulation was obtained for partial overlap (i.e., observed and performed action corresponded either in body side or movement kinematics), but not for full or no overlap conditions. The same pattern was obtained in Experiment 2 (back view instructions), thus ruling out a spatial compatibility explanation for the real-time pattern observed. Our results suggest that motor processes affect action prediction and real-time simulation. The strength of their impact varies as a function of the overlap between observed and executed actions.
The effect of mechanical transparency on adjustment to the visuo-motor transformation of a two-sided lever
Sandra Sülzenbrück & Herbert Heuer
For classic tools like a hammer or pliers, the relation between movements of the hand and the resulting movements of the effective part of the tool is mostly obvious and can be inferred from the mechanical features of these tools. This transformation between input and output variables is less transparent for modern types of tools like the computer mouse, where the location of hand movements is spatially separated from the observed effect of the tool. We conducted an experiment investigating the role of mechanical transparency for the acquisition of the visuo-motor transformation of a two-sided sliding lever. In this experiment participants controlled movements of a cursor on a monitor by moving the effort arm of a two-sided lever, with the position of the cursor representing the tip of the load arm of the same lever. One group of participants only saw the cursor on the monitor; for a second group, the cursor as well as the load arm of the lever was visible on the monitor. Group differences in movement times as well as in the accuracy of the internal model of the visuo-motor transformation will be discussed.
Anticipatory motor simulation of deceptive actions: psychophysics and TMS studies in soccer players
Enzo Tomeo & Cosimo Urgesi
Reaching high levels in sport performance implies not only motor activity but also the perceptive ability in predicting someone else?s actions. Research has shown the involvement of the motor system in the perception of actions. Here we used Transcranial Magnetic Stimulation (TMS) to record the activation of the corticospinal system in soccer players, goalkeepers and subjects unskilled in soccer during observation of videos of penalty kicks. In all subjects, the left primary motor cortex was stimulated while recording the EMG of the arm and leg muscles involved in the movements of kick and save. In a second psychophysical experiment, the three subject categories were asked to determine whether the observed penalty kicks would end to the right or the left of the goal. Video presentation could be interrupted before or after the beginning of the ball trajectory. Furthermore, videos could show typical or bluffing body actions, where discrepancy between body kinematics and the ball trajectory was inserted. Results demonstrated that the different experiences and motor ability of the three categories of subjects influenced the activation of the corticospinal system during action observation and their ability to foresee the fate of the kick.
The present results show that motor and visual expertise may exert a differential contribution to the development of the experts? abilities to predict the outcome of deceptive behaviors on the basis of the body kinematics and suggest that these differential abilities are reflected by different patterns of motor activation during action observation.
A Deep Policy Network
Daan Wierstra
We introduce the Deep Policy Network framework. Deep Policy Networks comprise a hierarchy of densely connected stochastic nodes that together constitute the policy of a reinforcement learning agent. Every node is a member of the exponential family, which includes, conveniently, continuous-valued distributions. Leave nodes are used for observation and action representation. Higher level nodes encode behavior and perception at more abstract levels. The deep structure combined with the use of multiple nodes to represent actions enable the framework to represent multimodal action strategies -- strategies where multiple highly different action composites are being evaluated at every point in time. This is especially useful in high-dimensional action domains such as robot arm control. The entire hierarchy learns by using the gradient on reward-modulated relative entropy differences in action distribution, and experiments confirm the viability of the approach.
Actions implied by domestic settings: The influence of spatial context on action recognition
Moritz Wurm
Many of our daily activities take place in rooms that are optimized for specific actions: In kitchens we prepare food, in bathrooms we engage in body care, but usually not vice versa. Hence, domestic rooms provide contextual information rendering some actions more likely to occur in a particular setting than others. This raises the question whether contextual information is exploited for the efficient analysis of actions. In analogy to the effects of spatial context on object recognition, we hypothesized that compatibility between contextual setting and action modulates neural activity, particularly in the motor system, during observation of ordinary everyday actions.
In an event-related fMRI experiment, subjects watched short videos of context-specific actions (e.g. cracking an egg, using a stapler) performed in rooms either compatible or incompatible with the action. Actions were also performed in empty rooms without any interior providing a neutral condition to distinguish between the effects of compatible versus incompatible settings.
We found increased activity in the left inferior prefrontal cortex (BA 44, 45, and 47) when actions were incompatible to the setting ? either compared to the compatible or the neutral setting. The parahippocampal place area (PPA) was activated when contrasting either of the room settings with the neutral setting.
Results indicate that contextual information provided by domestic settings modulates the perceptual analysis of actions. Two not necessarily mutually exclusive interpretations of inferior prefrontal activation can be discussed: Stronger neural responses for incompatibility could reflect either (a) interference of currently observed with context-implied actions, or (b) search and/or retrieval of higher-level action representations compatible with both the observed action and the setting.