1470 Part G Human-Centered and Life-Like Robotics and how that hand is moving.Other schemas implement 62.4.2 A Bayesian View hand-object spatial relation analysis and check how ob- of the Mirror System ject affordances relate to hand state.Together with F5 canonical neurons,this last schema(in parietal area 7b) We now show how to cast much that is known about provides the input to the F5 mirror neurons. the mirror system into a controller-predictor model(see In the MNS model,the hand state was defined as Miall et al.[62.54]and Wolpert et al.[62.1161)and a vector whose components represented the movement analyze the system in Bayesian terms.As shown by of the wrist relative to the location of the object and of the FARS model,the decision to initiate a particular the hand shape relative to the affordances of the object. grasping action is attained by the convergence in area Oztop and Arbib showed that an artificial neural net- F5 of several factors including contextual and object- work corresponding to PF and F5mimor could be trained related information;similarly many factors affect the to recognize the grasp type from the hand state tra- recognition of an action.All this depends on learn- jectory,with correct classification often being achieved ing both direct(from decision to executed action)and well before the hand reached the object,using activity inverse models (from observation of an action to acti- in the F5 canonical neurons that commands a grasp as vation of a motor command that could yield it).Similar training signal for recognizing it visually.Crucially,this procedures are well known in the computational motor training prepares the F5 mirror neurons to respond to control literature (see Jordan and Rumelhart [62.117] hand-object relational trajectories even when the hand and Kawato et al.[62.118]).Learning of the affordances is of the other rather than the self because the hand state of objects with respect to grasping can also be achieved is based on the view of movement of a hand relative to autonomously by learning from the consequences of the object,and thus only indirectly on the retinal input of applying many different actions to different parts of seeing hand and object,which can differ greatly between different objects. observation of self and other.Bonaiuto et al.[62.115] But how is the decision made to classify an ob- have developed MNS2,a new version of the MNS served behavior as an instance of one action or another? model to address data on audiovisual mirror neurons Many comparisons could be performed in parallel with that respond to the sight and sound of actions with char- the model for one action becoming predominantly ac- acteristic sounds such as paper tearing and nut cracking tivated.There are plausible implementations of this (see Kohler et al.[62.105]),and on the response of mir- mechanism using a gating network (see Demiris and ror neurons when the target object was recently visible Johnson [62.119]and Haruno et al.[62.1201).A gat- but is currently hidden (see Umiltd et al.[62.105]).Such ing network learns to partition an input space into learning models,and the data they address,make clear regions;for each region a different model can be ap- that mirror neurons are not restricted to recognition of plied or a set of models can be combined through an an innate set of actions but can be recruited to recognize appropriate weight function.The design of the gating and encode an expanding repertoire of novel actions. network can encourage collaboration between models The discussion of this section avoided any refer- (e.g.,linear combination of models)or competition ence to imitation(Sect.62.4.3).On the other hand,even (choosing only one model rather than a combination). without considering imitation,mirror neurons provide Oztop et al.[62.121]offer a similar approach to the esti- a new perspective for tackling the problem of robotic mation of the mental states of the observed actor,using Part perception by incorporating action (and motor informa- some additional circuitry involving the frontal cortex. tion)into a plausible recognition process.The role of We now offer a Bayesian view of using the predictor- 0 the fronto-parietal system in relating affordances,plans, controller formulation approach to the mirror system. 3 and actions shows the crucial role of motor information This Bayesian approach views affordances as priors in and embodiment.We argue that this holds lessons for the action recognition process where the evidence is con- neurorobotics:the richness of the motor system should veyed by the visual information of the hand,providing strongly influences what the robot can learn,proceeding the data for finding the posterior probabilities as mir- autonomously via a process of exploration of the envi-ror neurons-like responses which automatically activate ronment rather than overly relying on the intermediary for the most probable observed action.Recalling that of logic-like formalisms.When recognition exploits the the presence of a goal (at least in working memory)is ability to act,then the breadth of the action space be- needed to elicit mirror neuron responses in the macaque. comes crucially related to the precision,quality,and We believe it is also particularly important during the robustness of the robot's perception. ontogenesis of the human mirror system,for example.1470 Part G Human-Centered and Life-Like Robotics and how that hand is moving. Other schemas implement hand–object spatial relation analysis and check how ob￾ject affordances relate to hand state. Together with F5 canonical neurons, this last schema (in parietal area 7b) provides the input to the F5 mirror neurons. In the MNS model, the hand state was defined as a vector whose components represented the movement of the wrist relative to the location of the object and of the hand shape relative to the affordances of the object. Oztop and Arbib showed that an artificial neural net￾work corresponding to PF and F5mirror could be trained to recognize the grasp type from the hand state tra￾jectory, with correct classification often being achieved well before the hand reached the object, using activity in the F5 canonical neurons that commands a grasp as training signal for recognizing it visually. Crucially, this training prepares the F5 mirror neurons to respond to hand–object relational trajectories even when the hand is of the other rather than the self because the hand state is based on the view of movement of a hand relative to the object, and thus only indirectly on the retinal input of seeing hand and object, which can differ greatly between observation of self and other. Bonaiuto et al. [62.115] have developed MNS2, a new version of the MNS model to address data on audiovisual mirror neurons that respond to the sight and sound of actions with char￾acteristic sounds such as paper tearing and nut cracking (see Kohler et al. [62.105]), and on the response of mir￾ror neurons when the target object was recently visible but is currently hidden (see Umilta´ et al. [62.105]). Such learning models, and the data they address, make clear that mirror neurons are not restricted to recognition of an innate set of actions but can be recruited to recognize and encode an expanding repertoire of novel actions. The discussion of this section avoided any refer￾ence to imitation (Sect. 62.4.3). On the other hand, even without considering imitation, mirror neurons provide a new perspective for tackling the problem of robotic perception by incorporating action (and motor informa￾tion) into a plausible recognition process. The role of the fronto-parietal system in relating affordances, plans, and actions shows the crucial role of motor information and embodiment. We argue that this holds lessons for neurorobotics: the richness of the motor system should strongly influences what the robot can learn, proceeding autonomously via a process of exploration of the envi￾ronment rather than overly relying on the intermediary of logic-like formalisms. When recognition exploits the ability to act, then the breadth of the action space be￾comes crucially related to the precision, quality, and robustness of the robot’s perception. 62.4.2 A Bayesian View of the Mirror System We now show how to cast much that is known about the mirror system into a controller–predictor model (see Miall et al. [62.54] and Wolpert et al. [62.116]) and analyze the system in Bayesian terms. As shown by the FARS model, the decision to initiate a particular grasping action is attained by the convergence in area F5 of several factors including contextual and object￾related information; similarly many factors affect the recognition of an action. All this depends on learn￾ing both direct (from decision to executed action) and inverse models (from observation of an action to acti￾vation of a motor command that could yield it). Similar procedures are well known in the computational motor control literature (see Jordan and Rumelhart [62.117] and Kawato et al. [62.118]). Learning of the affordances of objects with respect to grasping can also be achieved autonomously by learning from the consequences of applying many different actions to different parts of different objects. But how is the decision made to classify an ob￾served behavior as an instance of one action or another? Many comparisons could be performed in parallel with the model for one action becoming predominantly ac￾tivated. There are plausible implementations of this mechanism using a gating network (see Demiris and Johnson [62.119] and Haruno et al. [62.120]). A gat￾ing network learns to partition an input space into regions; for each region a different model can be ap￾plied or a set of models can be combined through an appropriate weight function. The design of the gating network can encourage collaboration between models (e.g., linear combination of models) or competition (choosing only one model rather than a combination). Oztop et al. [62.121] offer a similar approach to the esti￾mation of the mental states of the observed actor, using some additional circuitry involving the frontal cortex. We now offer a Bayesian view of using the predictor– controller formulation approach to the mirror system. This Bayesian approach views affordances as priors in the action recognition process where the evidence is con￾veyed by the visual information of the hand, providing the data for finding the posterior probabilities as mir￾ror neurons-like responses which automatically activate for the most probable observed action. Recalling that the presence of a goal (at least in working memory) is needed to elicit mirror neuron responses in the macaque. We believe it is also particularly important during the ontogenesis of the human mirror system, for example, Part G 62.4