Learning for Sensorimotor Coordination in Robot Manipulation

This line of research aims to the production of autonomous intelligent systems with dexterous manipulation capabilities. The goal is the design and implementation of a full mobile manipulator system for service robotic tasks. Three main sublines can be distinguished:

1. Perception-Based Learning for Robot Manipulation Tasks under Uncertainty: Assumming that a solution for the symbol grounding problem is to use robotic capacities and particularly perception to ground the symbols in a symbol system, we approach Qualitative Spatial Reasoning from a bottom-up perspective: in what we call perception-based reasoning, the sensory information is represented at a subsymbolic level, the resulting category representations are then associated with some qualitative symbolic description of the input. We are dealing with complex robotic manipulation tasks under uncertainty involving contacts and the use of force/torque sensors. We use neural nets to learn to categorize sensory input: particularly Kohonen's self-organizing maps, incorporating new learning schemes. E. Cervera and A. P. del Pobil in cooperation with E. Marta and Prof. M. A. Serna (CEIT & Univ. de Navarra).
    
2. Intelligent Motion for a Robot Arm with Uncertainty Based on Computer Vision and Sensors: The aim of this project is the study, development and implementation of a fine-and-gross motion planning system for a robot manipulator for extraction, translation and insertion tasks. Its application in a flexible manufactoring cell is considered feasible. The system will lack a previous complete knowledge of the environment, which will be obtained from machine vision in the case of gross motion and force sensors in the case of fine motion. The 3D map obtained from stereo vision will be integrated into a previously developed motion planner, the start and goal placements will be identified too. Neural networks will be employed for error detection and recovery in fine motion. E. Cervera, P. Sanz, P.Martín, J. M. Iñesta, M. Pérez, B. Martínez and A. P. del Pobil in cooperation with F. Pla and M. A. López (Computer Vision Group, UJI).
    
3. Connectionist Reinforcement Learning Techniques for Motion Planning of Robot Manipulators: Analysis of reinforcement learning techniques along with the use of artificial neural networks and their application for acquiring goal-directed obstacle-avoidance reactive capabilities for sensor-based robot arms. This research aims to obtain reflexes that can allow a robot manipulator to reach a target position while avoiding collisions with the obstacles in an unknown environment. Learning of these capabilities relies only on range sensory information (e.g.: sonar sensors). Such reflexes are useful when the manipulator faces a completely new environment. P. Martín in cooperation with Dr. J. R. Millán (Institute for System Engineering and Informatics, European Commission Joint Research Centre, Italy).