Jump label

Service navigation

Main navigation

You are here:

Main content


Cognition is a central aspect of autonomously acting robots in so far as every decision and executed motion is ultimately based on knowledge and data supplied by sensors.

Information about the robot's own state (posture, spatial orientation, etc.) is mainly provided by interoceptive sensors such as accelerometers, gyroscopes and joint angle sensors. This sensor data needs to be fused and filtered appropriately, allowing the robot to move in a controlled and balanced way, and furthermore to interpret other sensor data properly.

Information about the robot's surroundings are gathered by exteroceptive sensors. Those may include distance sensors like sonar, infrared and laser scanners, but in our case mostly camera systems. To allow tracking of and reaction to fast movements, such information needs to be processed in real-time, meaning 20-30 times per second.


To achieve this on the computationally weak CPU of the standard platform league robots, the coloring and appearance of the soccer field is specified in advance. In previous years additional colored landmarks were placed around the field, but year by year those simplifications are removed to increase the challenge to the vision systems of the robots. Besides the removal of artificial landmarks and reshaping the goals to the form of real soccer goals, this also includes a loosening of the specification of the field's illumination.


Efficient Image Processing

The low-quality camera of the Aibo robots holds special problems for the processing of the captured images. A robust color segmentation is essential for the extraction of relevant objects. The Nao's cameras are of higher quality but the same restrictions generally still hold, especially with respect to the limited processing power available for the task.

Aibo-Bild Aibo-Bild_segm

colortable_raw colortable_gen

A generalization of the color classification based on radiation models can help to improve the robustness of the color segmentation in situations of variable illumination.

Additional advantages can be provided by sensor fusion. In the Standard Platform League three to four robots play soccer in one team. While playing autonomously the robots are also able to communicate via wireless LAN with each other. Therefore it is possible for one robot not only to access its own sensor information but also those of the other robots in its team to improve its own estimation about its surroundings.



In contrast to the limited cognition capabilities of the Aibo, when developing the humanoid robot Bender we decided to implement a catadioptric camera system (meaning it contains both light reflecting and refracting elements). This provides a nearly omni-directional field of view for the robot. The blind spot in front of the robot can be covered using an additional small conventional camera.

Bender_Kameras Bender_FOV arena5

Utilizing appropriate mathematical models of the image generation process it is possible to transform the resulting distorted omni-directional image into several perspective views or even a single panoramic view.

arena5_undist1 arena5_undist3 arena5_undist4 arena5_undist5


Such a transformation however is computationally expensive. To enable efficient image processing it is possible to incorporate the knowledge about the geometry underlying those catadioptric images into their direct processing. Combined with posture and orientation estimation based on measured joint angles, inertial sensors or information extracted from previous image frames the image processing can be done without expensive transformation in an efficient way and information about all relevant objects can be extracted in real-time.