• Sorted by Date • Classified by Publication Type • Sorted by First Author Last Name • Classified by Topic •
Çetin Meriçli and Manuela Veloso. Biped Walk Learning Through Playback and Corrective Demonstration. In AAAI 2010: Twenty-Fourth Conference on Artificial Intelligence, 2010.
Developing a robust closed-loop walking algorithm for a biped full-body humanoid robot is a challenging task due to the complex dynamics of the walk process in a high dimensional state and action space. While approaches that capture the analytical physical properties of the body have been proposed, they tradeoff the complexity of the model in favor of achieving tractability and real-time performance. Inevitably, the complications arising from the simplifications lead to a loss of balance while the robot is walking, leaving a difficult understanding on how to resolve the tradeoff between the complexity of an extended analytical model and its computational performance. In this paper, we contribute a two-phase biped walk learning approach, which is initially experimented on the Aldebaran Nao humanoid robot. In the first phase, we identify and save a "good walk cycle" from the motions of the robot while it is executing a given walk algorithm as a black box. We show how the robot can then play back such a recorded cycle in a loop to obtain a good open-loop walking behavior. In the second phase, we introduce an algorithm to directly modify the recorded walk cycle using real time corrective feedback provided by a human. The algorithm learns joint movement corrections to the open-loop walk based on the corrective feedback as well as the robot's sensory readings while walking autonomously. Compared to the open-loop algorithm and hand-tuned closed-loop walking algorithms, our two-phase method provides a significant improvement in walking stability, as demonstrated by our experimental results.
@inproceedings{mericli2010c,
author = {Çetin Meriçli and Manuela Veloso},
title = {Biped Walk Learning Through Playback and Corrective Demonstration},
booktitle = {AAAI 2010: Twenty-Fourth Conference on Artificial Intelligence},
year = {2010},
abstract = {Developing a robust closed-loop walking algorithm for a biped full-body humanoid robot is a challenging task due to the complex dynamics of the walk process in a high dimensional state and action space. While approaches that capture the analytical physical properties of the body have been proposed, they tradeoff the complexity of the model in favor of achieving tractability and real-time performance. Inevitably, the complications arising from the simplifications lead to a loss of balance while the robot is walking, leaving a difficult understanding on how to resolve the tradeoff between the complexity of an extended analytical model and its computational performance. In this paper, we contribute a two-phase biped walk learning approach, which is initially experimented on the Aldebaran Nao humanoid robot. In the first phase, we identify and save a "good walk cycle" from the motions of the robot while it is executing a given walk algorithm as a black box. We show how the robot can then play back
such a recorded cycle in a loop to obtain a good open-loop walking behavior. In the second phase, we introduce an algorithm to directly modify the recorded walk cycle using real time corrective feedback provided by a human. The algorithm learns joint movement corrections to the open-loop walk based on the corrective feedback as well as the robot's sensory readings while walking autonomously. Compared to the open-loop algorithm and hand-tuned closed-loop walking algorithms, our two-phase method provides a significant improvement in walking stability, as demonstrated by our experimental results.}
bib2html_pubtype = {Refereed Conference},
bib2html_rescat = {Learning from Demonstration},
bib2html_dl_pdf = {../files/cmericliAAAI2010WalkLearning.pdf},
}
Generated by bib2html.pl (written by Patrick Riley ) on Tue Nov 07, 2017 00:34:04