Contributed by: Wolfram Schenck , Bielefeld University

Introduction
The notion "cognitive robotics" (CR) has two different meanings: On the one hand, CR refers to a research method within cognitive science, on the other hand, to a research direction within robotics. In practice, many CR projects are related to both aspects.

Cognitive robotics as research method
The subject of cognitive science are cognitive systems. The research goal is to explain the mechanisms of information processing in natural cognitive systems (like the human mind) and to replicate them in artificial agents. To approach this goal, two basic research methods are available: analytical experiments and synthetic modeling. Both methods require models and theories about the human mind. In the analytical approach, these models are used to specify experimental hypotheses which are tested on human subjects in behavioral or neurophysiological studies. In synthetic modeling, these models are implemented on artificial agents to test if they are actually complete and powerful enough to enable the hypothesized cognitive abilities. For synthetic modeling, one can use both simulated agents and real-world agents—thus robots. In summary: CR is the field in cognitive science which applies robots for synthetic modeling.

The robotics approach to cognitive science offers some genuine advantages for the process of model testing. Webb (2000) writes on p. 546 with regard to the related field of biorobotics: "Robots as models are a means by which hypotheses can be tested for adequacy and sufficiency to explain a set of data, and additional predictions from the hypotheses can be derived." Moreover, Webb (2000) emphasizes that the use of robots enforces the researcher to characterize the problem thoroughly and to consider and understand the role of the environment. In a pure computer simulation, the simulated environment of the agent can lack important properties of the real physical world with regard to the tested model (see also Grasso, 2001). By using robots, this risk is reduced (although not completely eliminated; e.g., the sensor equipment of the robot may be insufficient, causing misleading results). In addition, robot models enforce completeness: It is not possible to omit any part of the sensorimotor loop — the tested model works only in its full implementation, including the sensory and the motor part (and the integrated processing in between). This enhances the validity of the results, helps to identify missing or wrong parts of any model, and facilitates the generation of new hypotheses.

The use of robots in cognitive science fits very well to the research paradigm of "new artificial intelligence" (Pfeifer and Scheier, 1999). Its basic claim is that intelligent behavior can only emerge in agents which are embodied, situated, and adaptive. Such agents have a real physical body with a specific morphology and specific sensory and motor capabilities, they act in a specific environment, and they are able to learn from experience. New AI states that high-level cognition emerges from the sensorimotor relationships an agent’s body experiences during the interaction with its environment. Thus, cognitive abilities can only be explained with reference to the body and the environment of an agent.

Following the premises of embodiment, situatedness, and adaptivity, research in CR focusses mainly on models of sensorimotor coordination as the (future) building blocks of highlevel cognition. Important general research questions concern how sensorimotor models can be learned, and how they can be used to guide behavior and to facilitate perception and cognition. For example, the "Darwin" series of mobile robots (Almassy et al., 1998; Krichmar and Edelman, 2002) were developed to show by synthetic neural modeling how certain perceptual skills emerge by the interaction of the robot with its environment. On both Darwin V (Almassy et al., 1998) and Darwin VII (Krichmar and Edelman, 2002) a rather detailed model of the brain structures which are involved in visual processing and visuomotor coordination was implemented. During the interaction with the environment, this synthetic brain developed through changes in the synaptic strengths. In this process, specialized neural structures emerged which are linked to perceptual skills like invariant object recognition or experience-dependent perceptual categorization. The authors claimed that by exhaustive analysis and manipulation of these artificial brain structures, this approach provides valuable heuristics for understanding the interactions in the real brain.

A second example for cognitive robotics research is the recent study by Bongard et al. (2006). They proposed an active self-modeling process for a four-legged robot. In this process, the internal sensorimotor self-model of the robot is first generated by model synthesis through directed exploration. Explorative movements are not generated at random, but chosen according to which movement would be the most informative for the identification of the self-model. Afterwards, the self-model is used for action generation until an unexpected sensorimotor pattern is detected. In this case, the self-modeling process starts again. The authors showed that the system is even able to cope with the loss of one leg. They concluded that their "work suggests that directed exploration for acquisition of predictive self-models may play a critical role in achieving higher levels of machine cognition" (Bongard et al., 2006, p. 1121).

Cognitive robotics as research direction
Robotics has a wide spectrum of applications and research directions. "Classical" robotics aims on designing industrial robots as sophisticated automata for special tasks (like the assembly of cars). The underlying research focusses mainly on mechatronics, electronics, and control theory. In contrast, the main goal of CR is the construction of autonomous robots which act in an "intelligent" way. In this context, the term "intelligence" does not refer to skilled analytical thinking, but more to "natural intelligence" as it is also exhibited by many animals. To illustrate this point: A cognitive robot should be able to act autonomously in the real world in its specific "ecological niche" under a wide range of environmental conditions, it should have a non-trivial repertoire of actions to choose from, and it should be able to learn from its experience. Moreover, many possible applications of cognitive robots would require them to communicate and interact with humans in a complex way.

The construction of cognitive robots is not necessarily linked to any specific research paradigm, thus cognitive architectures which are based on symbol processing can be combined with adaptive sensorimotor models. Furthermore, CR research focusses on speech and dialog systems, computer vision, and human-robot-Interaction (see for example the COGNIRON project: http://www.cogniron.org).

References

• N. Almassy, G. M. Edelman, and O. Sporns. Behavioral constraints in the development of neuronal properties: A cortical model embedded in a real-world device. Cerebral Cortex, 8 (4):346–361, 1998.
• J. Bongard, V. Zykov, and H. Lipson. Resilient machines through continuous self-modeling. Science, 314(5802):1118–1121, 2006.
• F. W. Grasso. Environmental information, animal behavior, and biorobot design. Reflections on locating chemical sources in marine environments. In B. Webb and T. R. Consi, editors, Biorobotics, pages 21–35. MIT Press, Cambridge, MA, 2001.
• J. L. Krichmar and G. M. Edelman. Machine psychology: Autonomous behavior, perceptual categorization and conditioning in a brain-based device. Cerebral Cortex, 12(8):818–830, 2002.
• M. Lungarella, G. Metta, R. Pfeifer, and G. Sandini. Developmental robotics: a survey. Connection Science, 15(4):151–190, 2003.
• R. Pfeifer and C. Scheier. Understanding Intelligence. MIT Press, Cambridge, MA, 1999.
• B. Webb. What does robotics offer animal behaviour. Animal behaviour, 60(5):545–558, 2000.