Paper Info
Title
A scalable method for multi-stage developmental learning for reaching.
Abstract
In this paper, we introduce a technique to learn sensory-motor sequences in multiple consecutive stages, where one stage bootstraps sequences serving as training data for the subsequent stage. By introducing multiple interaction stages and recording the generated sensory-motor sequences of a preceding interaction stage, we obtain a system capable of self-generation of training data to increase the skill performance over time. At the beginning, our system uses a constrained degrees of freedom (DOF) exploration to gather a simple and short set of training data for a meaningful first-stage behavior. This minimum amount of samples already enables the robot to generate a reaching behavior for goals in the visual field. The generated observed sensory-motor sequences are then used as training data for a subsequent reaching phase. We are using the Predictive Action Selector (PAS) as a system building block, which provides bootstrapping of visual-proprioceptive predictions. Since our system was already presented on a robot with 2 DOF and 5 DOF, we proceed with the evaluation on a different robot with 6 DOF. Thus, we demonstrate the generality of the approach on various robotic platforms with different morphologies. By increasing the number of DOF, we continue showing the scalability of the presented system. Without any prior knowledge of neither the forward nor the inverse kinematics, the experiments show promising results with a reaching success rate of 66% during the first-stage reaching. This result is obtained by using only 13 training sequences (349 samples) which have been obtained during the constrained DOF exploration in only a few minutes. The developmental process is then shown by taking the generated sequences obtained during the first-stage reaching and using them as training data for the second-stage reaching. With second-stage reaching, the goal reaching times were reduced by up to 59% and, in contrast to first-stage reaching, it allows continuous retraining with increasing training data in subsequent stages.
Year
Venue
Field
2017
Joint IEEE International Conference on Development and Learning and Epigenetic Robotics ICDL-EpiRob
Computer vision,Inverse kinematics,Visualization,Bootstrapping,Computer science,Robot kinematics,Artificial intelligence,Robot,Retraining,Generality,Scalability
DocType
ISSN
Citations 
Conference
2161-9484
2
PageRank 
References 
Authors
0.37
0
4
Name
Order
Citations
PageRank
Wolfgang Burger171.15
Erhard Wieser2102.31
Emmanuel C. Dean-Leon36215.39
Gordon Cheng41250115.33