Information processing within autonomous robots should follow a biomimetic approach. In contrast to traditional approaches that make intensive use of accurate measurements, numerical models and control theory, the proposed biomimetic approach favors the concepts of perception, situation, skill and behavior - concepts that are used to describe human and animal behavior as well. Sensing should primarily be based on those senses that have proved their effectiveness in nature, such as vision, tactile sensing and hearing. Furthermore, human-robot communication should mimic dialogues between humans. It should be situation-dependent, multimodal and primarily based on spoken natural language and gestures. Applying these biomimetic concepts to the design of our robots led to adaptable, dependable and human-friendly behavior, which was proved in several short- and long-term experiments

To fit well the needs of autonomous mobile manipulator, two robust adaptive Neuro-Fuzzy motion controllers are developed. The first controller, based on a computational efficient processing scheme for fuzzy reactive navigation, is used to generate the commands for the servo-systems of robot arm so that, locally, it may choose its way to its goal autonomously. The second fuzzy reactive navigation is implemented in mobile platform so that it maintains a permanent flexible path between two nodes in network generated by a probabilistic roadmap approach. In order to consider the compatibility of stabilisation, mobilisation and manipulation, we derive a coordinated fuzzy local planner algorithm so that the mobile manipulator can avoid stably unknown and/or dynamic obstacles The purpose of an integration of robust controller and Modified Elman Neural Network is to deal with unmodeled bounded disturbances and/or unstructured unmodeled dynamics.

Robots that never need any calibration of their kinematics, their actuators or their sensors promise great advantages in terms of maintenance cost, robustness and adaptability. Here we propose an approach for calibration-free vision based navigation in networks of corridors. The underlying concepts are based on the utilization of laws of projective geometry that always apply, regardless of camera characteristics, and in some cases on learning by doing. Our approach is based on a vehicle-mounted camera looking into an approximately forward direction as the main sensor and a qualitative knowledge of the appearance of objects of interest, such as floors and walls of corridors. The skills developed on the basis of the approach were intensively tested and have proved their effectiveness, adaptability and robustness under real-world conditions in two different office environments and with two robots, a humanoid personal robotic assistant, HERMES and a transportation robot, ATHENE II.

We have developed a humanoid robot, HERMES, to study several key technologies that are important for personal robots, such as robot design, sensors and perception, locomotion, localization and navigation, manipulation, human-robot communication and interaction, adaptability and learning, system architecture and integration, and dependability. The robot's skill-based system architecture was derived from a qualitative model of human information processing and insights gained from psychological literature dealing with skill acquisition, human performance and motor learning. HERMES' system architecture, several of its skills and the design principles are introduced, and some experiments carried out with the real robot are presented, including a long-term test where HERMES served in a museum, far away from its home laboratory, for more than six months up to 12 hours per day. During this period the robot and its skills were regularly demonstrated to the public by nonexpert presenters. Also, HERMES interacted with the visitors, chatted with them in English, French and German, answered questions and performed services as requested by them.

An approach to realize learning calibration-free stereo vision-based robot for manipulating objects is introduced. It allows a robot gather experiences through interaction with the world and continuously improve its performance based on the collected experiences. It uses a direct transition from image coordinates to motor control commands, but no world coordinates and no inverse perspective or kinematic transformations. The approach has been tested in real-world experiences on an uncalibrated vision-guided manipulator with five degrees of freedom to grasp a variety of differently shaped objects.

An approach to control vision-guided robots is introduced. It allows searching and grasping differently shaped objects that may be located anywhere in the robot's work space, even not visible in the initial fields of view of cameras. It eliminates the need for a calibration of the robot and of the vision system, it uses no world coordinates and no inverse perspective or kinematic transformations, and it comprises an automatic adaptation to changing parameters. The approach has been implemented on a calibration-free vision-guided manipulator with five degrees of freedom (DOF) and was evaluated in real-word experiments.

A novel concept for recognition and classification of differently formed objects is introduced. It can be used for manipulating objects by uncalibrated vision-guided manipulators. The object recognition is achieved according to an object-oriented and model-based approach. The task is decomposed into several largely independent modules, and each module is responsible for a limited subtask. The emphasis in this article is on the object classification. It is achieved by observing directly the object appearance in real-time images without precise quantitative knowledge about the manipulated objects. The approach was evaluated and demonstrated in real-word experiments on an uncalibrated vision-guided manipulator with five degrees of freedom for recognition, classification and manipulation of a variety of differently shaped objects in nearly arbitrary orientations and positions anywhere in the robot's 3D work space.

In vision-based robot control typically more sensor data are available than the minimum necessary for computing robot control commands. We propose an approach for utilizing the redundant sensor data to improve the robustness of robot control. It is based on a novel method for taking into account all available measurements in solving the resulting over-determined system of equations for the robot control commands. The approach has been evaluated in computer simulations and in real-world manipulation experiments with a calibration-free robot. It has proven to be suitable for a real-time implementation and to lead to a more robust control than conventional methods.

In vision-based robot control typically more sensor data are available than the minimum necessary for computing robot control commands. We propose an approach for utilizing the redundant sensor data to improve the robustness of robot control. The key point is that the sensor data space is transformed into a space of lower dimensionality before computing the robot control commands, thus avoiding to solve an over-determined system of equations. Compared to the rather simple alternative method of truncating measurements it has the advantage that, in the interest of robustness, it takes into account all available measurements. The approach has been evaluated in real-world manipulation experiments with a calibration-free robot. It has proven to be suitable for a real-time implementation and has led to a more robust control than the alternative approach.

An efficient approach to object recognition was developed. It eliminates by a model-based subsampling of images the differences of the sizes of object images that are caused by varying distances. The distance-invariant images both simplify and-due to their reduced number of pixels-help to accelerate object recognition. This model-based subsampling has been used for creating a database of distance-independent representations of various objects occurring in natural scenes allowing the subsequent recognition of such objects in real time. An interactive user interface with a learning ability was provided to facilitate the introduction of new objects into the database. A number of algorithms for recognizing objects were implemented and evaluated. They employ different forms of object representations and were analyzed regarding their effectiveness for recognizing objects in varying distances. In experiments two of the investigated recognition methods, one based on cross correlation and the other one on user-defined edges, appeared suitable for realizing a fairly reliable object recognition in real time, as required by mobile robots operating in natural terrain.

As part of a driver support system for motor vehicles on freeways, an obstacle recognition system was developed within the EUREKA project PROMETHEUS over the last four years. The obstacle recognition system uses a single monochrome TV camera and a multi-processor robot vision system. The recognition system includes a road tracker and an obstacle detector that were described elsewhere, and an object classification and tracking module presented here. The module is based on generic 2D object models. It recognizes obstacles in real time and from distances of 200 to 300 m, sufficient for high-speed driving. The module was extensively tested in real-world scenes on the German Autobahn and on various other roads, including city streets with dense traffic.>

Taking an autonomous road vehicle as an example, approaches and methods for object detection (this being the first step of obstacle recognition) are discussed. Vision, using a single passive camera as the primary sensor and a real-time vision system for analyzing the resulting image sequence, is found to be the best approach. Results obtained in real-world experiments with such a system are reported: large obstacles have been detected in distances up to 716 m; all relevant objects that were encountered during a prolonged test on a highway were detected; the cycle time of the detection process was below 100 ms, and less than one false alarm occurred per minute.

A mobile robot operating in the presence of other vehicles should be able to determine at any time whether some other vehicle is in the process of passing it. A concept of a recognition system for realizing this ability is introduced. An algorithm is presented that interprets the image sequence of a forward-looking camera and detects passing vehicles when they enter the field of view. Controlled correlation is used for feature extraction, and the rate of false detections is minimized by applying a motion criterion. The algorithm operates in real time. It has been tested with recorded highway scenes played back and interpreted in real time.>