Real world applications are usually subject to change and very difficult to be predicted. Any sudden changes in the environment can possibly cause temporary lose in communication with the external world. Some organisms, those that have the ability of cognition or thinking, can cope with such situations by replacing the external missing or corrupted sensory data with their own internal representation (or experience).

In recent decades, a branch of science called cognitive neuroscience, an interdisciplinary link between cognitive psychology and neuroscience, has been established to introduce such phenomena to the mobile robot [1]. It was hoped that adding this feature to the robot would move autonomous robots closer to interfacing with real world applications.

Cognitive Roboticsis concerned with endowing robots with mammalian and human-like cognitive capabilities to enable them to accomplish complex tasks in complex environments. Cognitive ability is the ability to understand and try to make sense of the world. In [2], the authors have argued that all living creatures are cognitive to some degree. Several authors have argued in recent years that cognition and consciousness can be achieved to some extent on the mobile robots [3-5]. We believe that the level of or how much the robot could be conscious of the surrounding environment depends on how much the robot knows about this environment. Cognition in robots includes perception processing, attention allocation, anticipation, etc. One of the possible approaches to measuring these capabilities in the robot is by examining its ability to cope with missing external sensory data during performance of a specific task. Said in a different and operational way, it is the ability of a robot to perform blindfolded navigation, where the robot navigates within a known environment using only its internal representation.

In recent years, building a complete blindfolded navigation system in a mobile robot has been a challenging task for many robotic researchers [6-9]. For instance, some initial experiments were presented in [7] that aim to contribute toward building a robot that navigates completely blindfolded in a simple environment using a two-level network architecture; 1) low-level abstraction from sensorimotor values to a limited number of simple abstract “concepts”, following the work done by Linker and Niklasson [6,10], and 2) higher-level prediction/ representation of the agent’s interaction with the environment, inspired by the work done by Nolfi and Tani [11]. These efforts have to some degree succeeded in allowing the robot to anticipate long chains of future situations. However, they have failed to support a completely blindfolded navigation [8], in which the robot repeatedly uses its own internal representation values instead of the real sensory inputs for a certain number of times for its navigation. The failure partly seems to be due to the short range of the robot’s proximity sensors that they used, which limits the amount of data that could be abstracted from the environment. The consequence of this limitation is that the robot does not have enough sensitivity about the environment. We argue here that improving the robot’s sensorimotor abstraction level, therefore, could possibly overcome this problem. For instance, instead of relying only on the limited data provided by the robot proximity sensors, let the robot see the environment using its camera, abstract enough data, and arrange it well in its memory to aid in building its internal representation.

To support our argument, we have done psychological experiments, similar to the one introduced by Lee and Thompson [12] with a little change. In a series of two experiments, we demonstrated the accuracy with which humans can guide their behavior based only on their internally sensory experiences. Two subjects were asked to do the same task under different conditions. The first subject X was asked to “look” around in a given room and locate a specific target (Figure 1(a)). He was then blindfolded and asked to locate the target again. The subject performed the task accurately with closed eyes, in the same manner as when he was free to “look” (Figure 1(b)). However, he could not predict the exact time needed to turn to the target and this caused the two hits with the obstacle (the empty circles in Figure 1(b)). The second subject Y was not allowed to explore the room with his eyes (no vision input). Instead, he was blindfolded and walked around the room touching things around him until he found the target (Figure 1(c)). He was then asked to seek the target again blindfolded from the initial position. Though successful in reaching the target, he took more time than that needed by subject X. In addition, the number of times that he hit the wall or touched it to correct or locate his direction was greater (Figure 1(d)).

From the above experiment we can conclude that subject X had collected a sufficient amount of data from the environment during his first “eyes open” navigation. This data could be various dimensions in the room which the

subject related to times and distances that helped him to build internally—in his inner world where sensory experiences and consequences of different behaviors may be anticipated—his own internal image. In contrast, the amount of data that subject (Y) had collected was limited to the objects that his hand touched during his first blindfolded navigation and their relation to his moving steps. This data, however, was not good enough to accurately perform the task.

In the above experiment, subject Y could be a demonstration of the results of the most recently reported works (e.g., [7]), since they used the short-range proximity sensors for building the sensorimotor abstraction level.

We also tried to demonstrate the inner world that was automatically built inside both subjects’ memory by giving each of them a sheet of white paper and asking them to draw the outline of the room that they trained in (note that subject Y had never seen the room). It was not surprising to find out that subject X could draw almost all the details of the room (Figure 2(a)). However, subject Y could hardly draw the layout of just the objects that he touched during his movement (Figure 2(b)).

The work presented in this paper was motivated by the problems described above. Here we explore the inner world of a real mobile robot that has a chance to explore the surrounding environment with its camera before it was told to navigate blindfolded in it. In this study, the robot used two network architectures. The first was used to control its navigation, while the second, to build its internal representation.