Novel approach to stereo vision and calibration-free robot control:

• Direct transition from sensor data to motor control commands
• No world coordinate system
• No calibration of camera or arm

Relationships between motor control commands and resulting changes in the images ("Sensor-Control Jacobian") are learned automatically:

• Either through test motions, or as side effects of manipulation processes
• No training phase necessary
• Automatic correction or forgetting of obsolete knowledge

Results:

• Robustness against model errors and environmental changes
• Adaptability regarding even severe disturbances of the sensor configuration, changes in the environment or aging of components
• Automatic improvement of manipulation speed and skill ("learning by doing")