The group research is focused on the paradigm of sensing, representation and knowledge organization for decision making in complex and uncertain domains, dealing with the design and implementation of reactive/cognitive architectures. Such models are applied for decision making under incomplete, imprecise and uncertain knowledge on the system-environment in different fields of the human activity such as: Autonomous mobile robots indoor (Offices, Parking, Warehouses), outdoor (Agriculture, Horticulture, Forestry), Quality Control.

A schematic of the research steps from perception to action :

• Patterns discovery, classification (bayesian, neural nets, fuzzy) and learning
• Sensor integration and active perception
• Decision making architectures in complex and uncertain domains
• The architectures ESTOPA and AMARA
• Applications:
• Autonomous mobile robots
• Semi-autonomous inspection for quality control

During last ten years several physical architectures have been designed, implemented and evaluated for Multi-sensor Integration and Artificial Perception. From initial single processor architectures, multiprocessors dedicated networks (T800, C40) are proposed to solve ever and ever increasing demanding system performances. At present time TCP/IP subnet are links between P6 nodes for early stages development and cost performance optimization. Here pheripheral information interchange is performed by autonomous modules based on last generation microcontrollers. In this direction micro-controllers display obvious complementary performances to the conventional distributed processing nets offering high communication capabilities with real environments for information acquisition or controlling actions, but limited computing and memory resources. The modularity, connectivity and power processing that offers such frames allows for a systematic and incremental approach to complex systems.

On this distributed architectures gravitates all information analysis and synthesis algorithms grounded on Analytic Models. Behavior modification based on self-experience (empirical learning) for statistical models generation concerning both the environment and the own system capabilities, has give rise to the ESTOPA Architecture. This architecture is devoted to the automatic features selection, competition among classifiers and definition of the global interpretation and evaluation criteria concerning the Models Recognition performance.

Schemes for Data Fusion and empirical learning have been developed pointing to an Active Perception for artificial behavior generation in the fields of machining processes, inter-modal transport monitoring, and autonomous mobile robots. On the other hand, the complexity and uncertainty inherent to real systems has focused our attention to more flexible representation schemes to handle knowledge. Thus, taking into account that much of human knowledge useful in control processes is expressed by means of linguistic inference sentences, Heuristic Models are proposed to manage uncertainty in sensing, modeling and decision making in the system-environment relationship.

Active Perception allude to the internal and external reconfiguration processes that hold in a system for purposive goals accomplishment. In such a perceptual framework all components of the perceptual structure can dynamically participate, deciding the system relevant components as: suited sensors, dynamic range, the signal processing spectral window, the net connections topology, the processes distribution, the descriptors space, the learning mechanisms, or the criteria for the evaluation of the results. In systems with a high complexity the searching spaces may become intractable directing us to an heuristic segmentation process to perform local search by using either Fuzzy Knowledge Bases or Genetic Algorithms.

In collaboration with:

• The Optics Institute-CSIC in Madrid and within the Active Vision paradigm, the study of local representation of images has been proposed to obtain functional operators models able to deal with intermediate and primary vision tasks. Its final goal would be the integration of fundamental aspects from computational neuroscience, image processing and pattern recognition. A key issue is the information representation models on the images (multi-resolutional, selective attention, foveatization, neural or connectionist)
• The Microelectronics National Center-CSIC in Sevilla, we are facing the implementation of VLSI circuits in the design and construction of an Artificial Retina for primary vision tasks. This retina will be used in two already funded projects dealing with:
• Outdoor robots navigation with a multi-sensor integration architecture
• Defects detection in quality control

The perceptual roots of our group, have enabled an integrative viewpoint of the different competing architectures (multi-resolutional, subsumed, behavior-based, tasks control). This is way since 1992 an Hierarchical Multi-Agents Architecture (AMARA) has been proposed, as a flexible and incremental framework able to deal with:

• Complexity by means of a levels structure that cut the possibilities tree
• Dynamic environments and sensors failures through a set of reactive and well fitted Basic Agents at the lowest level of the architecture

In this proposal the efficiency concept, as a measure of the adequateness between the system structure and the required functionality, joints the concept of evolution to make a parallelism between the artificial schemes (or Agents of Behavior) and the behavior generation mechanisms in human being. The need of good knowledge representation as a way to pursue intelligence resume the paradigm that will be explored and investigated in deep in the next years.