-We introduce a simple example of artificial system which aims to mimic the process of the mirror self-recognition ability limited to very few species. We assume that evolution of the species is reflected in the structure of the underlying control mechanism and design modules of the mechanism, concerning its incremental development. On this example, we also demonstrate modular architecture suitable for such task. It is based on decentralization and massive parallelism and enables incremental building of control system which is running in real-time and easily combines modules operating at different pace.
Seeing in the mirror, we recognize ourselves. However
child less than eighteen months old is rather looking for
another person behind the mirror. In nature, the mirror
selfrecognition is present only in exceptional cases, for instance
chimpanzees recognize themselves in the mirror, while cats do
not. Can a robot recognize itself in the mirror? And what is the
origin of the self-recognition ability? We follow Scassellati and
Hart [
Using the simulator of the humanoid robot iCub [
For examination of mechanisms underlying the mirror
selfrecognition process, we employ a modular framework
(AgentSpace Architecture [
Now we can try to implement the system, using which we are able to evaluate correlation between own body movement and the seen image movement. We aim to implement a working demo following biological relevancy of our approach, based on knowledge about the above-mentioned species.
The body model is provided by the iCubSim simulator, thus we need to implement just its control and monitoring from the blackboard. We implement a module motor which reads an intended joint position intention from the blackboard, communicates with the simulator (via yarp rpc protocol) and as a result, it writes proprioception to the blackboard which represents a real joint position (Fig. 1). In parallel, the added camera regularly provides a new image to the blackboard.
As we aim to compare the body model (i.e. value proprioception) with the image captured by the camera, we need to get an analogical model from the image. This approach is still biologically relevant since several species are proven to obtain such ability (mirror neurons). (How such ability has emerged and evolved, is not the subject of our exploration – it is only limited to the implementation of this ability.) Since we take into account both interaction between the robot and human and interaction between the robot and its image in the mirror, we combine two specific methods which output the same data onto the backboard – the model of the seen image (the model value in Fig. 1).
For the iCubSim image processing (we rotate a printed picture of iCubSim in front of the camera) we employ the SURF method. SURF provides projection of given template to the image seen from which we can easily calculate inclination of the template. For the processing of human image (interacting person is moving his head in front of the camera) we use combination of Haar face detector and CamShift tracking: the Haar detector detects face in the upright position and sets up template for CamShift to track. The projection of the template to the seen image calculated by CamShift provides us with the required head inclination. All algorithms employed are available in OpenCV library (www.opencv.org).
The correlation of these two comparable models can only be evaluated by their changes through time therefore both robot and human have to be put in motion. One of the mechanisms, that can provide it, is imitation. Due to the utilized mirroring, a passive imitation can be implemented by direct copying of values of the body model seen (i.e. the model value) into the robot’s body model (i.e. the intention value). As a result, iCubSim imitates side to side movement of one’s head when moving in front of the camera. Optionally we can also implement the active part of imitation, which can be called the invitation to imitation and obtains a higher priority than the passive imitation and occasionally generates side to side movement of the robot’s head when the body seen does not move. However we found out that even slight inaccuracies in the model evaluation are sufficient to induce the imitation process.
Imitation process provides us with the data of correlation between the body owned and the body seen. When an individual lives within a society, it is important for him to categorize the data and associate them with the image seen. When such individual meets the mirror, it possibly creates a new category and finds out that the corresponding correlation is unusually high. In fact, it is so high that the image seen can not only be associated with the body model seen, but also with its own body, meaning that it sees itself. Instead of modelling the society, for our purposes, it is sufficient to record the data produced by the two models since the aim of the study is to differentiate the data created when robot encounters human and when robot sees its own image in the mirror.
As a result of implementation of the above mentioned mechanisms, when iCubSim's image is reflected into the external camera by a mirror, robot invites himself to imitation and it stays in an interaction with itself for certain amount of time (Fig. 2). Thus correlation between the body model (proprioception) and the model of the image seen (model) can be evaluated and its unusually high value indicates that the robot sees itself in the mirror. Thus the designed structure of the mirror self-recognition process is partially verified.
http://www.agentspace.org/mirror/iCubSimAtTheMirror.mp4