|
Company Name: AG Technische Informatik -
Company Type: Research & Development
Company Category:
Robotics Instruments
Company Profile
Based at the University of Bielefeld. Main research topics include biorobotics, cognitive robotics, sensorimotor control, and parallel computation.
About Company
Adaptive motor learning faces several problems. First, there is the problem of the missing teacher signal. While the sensory error after a failed motor action is known (e.g., the ball has been thrown too far), the correct motor command which is needed for controller training remains unknown. Second, the sensorimotor space (comprising all sensory and motor dimensions which serve as necessary inputs and outputs for the motor controller) is most times of very high dimensionality, making it infeasible to search at random for the right motor commands. And third, motor tasks often imply a one-to-many mapping. For example, it is often possible to use different arm postures for pointing to a certain target (at least elbow-up and elbow-down). This means that for a certain input of the motor controller (the position of the pointing target) many different valid outputs (motor commands) exist. This has severe implications for the implementation of adaptive motor controllers in robotic applications: All neural network techniques which rely on function approximation are useless for this purpose, but instead advanced neural network architectures and machine learning techniques have to be developed.
|
