Learning manipulation skills by combining PCA technique and adaptive interpolation

We propose an approach for learning manipulation skills of robot arms based on complex sensor data. The first component of the used model serves as a projection of high-dimensional input into an eigenspace by using Principal Component Analysis (PCA). Complex sensor data can be efficiently compressed if robot movements achieving optimal manipulation tasks are constrained to a local scenario. The second component is an adaptive B-spline model whose input space is the eigenspace and whose outputs are robot motion parameters. In the offline learning phase, an appropriate eigenspace can be built by extracting eigenvectors from a sequence of sampling sensor patterns. The B-spline model is then trained for smooth and correct interpolation. In the online application phase, through the cascaded two components, a sensor pattern can be mapped into robot action for performing the specified task. This approach makes tasks such as visually guided positioning much easier to implement. Instead of undergoing cumbersome hand-eye calibration processes, our system is trained in a supervised learning procedure using systematical perturbation motion around the optimal manipulation pose. If more sensors or some robust geometric features from the image processing are available, they can also be added to the input vector. Therefore, the proposed model can integrate multiple sensors and multiple modalities. Our experimental setup is a two-arm robotic system with "self-viewing" hand-eyes and force/torque sensors mounted on each parallel jaw gripper. Implementations with one-hand grasping and two-hand assembly based on visual and force sensors show that the method works even when no robust geometric features can be extracted from the sensor pattern.