Adversarial training is not ready for robot learning
Lechner M, Hasani R, Grosu R, Rus D, Henzinger TA. 2021. Adversarial training is not ready for robot learning. 2021 IEEE International Conference on Robotics and Automation. ICRA: International Conference on Robotics and AutomationICRA, 4140–4147.
Download (ext.)
Conference Paper
| Published
| English
Author
Department
Grant
Abstract
Adversarial training is an effective method to train deep learning models that are resilient to norm-bounded perturbations, with the cost of nominal performance drop. While adversarial training appears to enhance the robustness and safety of a deep model deployed in open-world decision-critical applications, counterintuitively, it induces undesired behaviors in robot learning settings. In this paper, we show theoretically and experimentally that neural controllers obtained via adversarial training are subjected to three types of defects, namely transient, systematic, and conditional errors. We first generalize adversarial training to a safety-domain optimization scheme allowing for more generic specifications. We then prove that such a learning process tends to cause certain error profiles. We support our theoretical results by a thorough experimental safety analysis in a robot-learning task. Our results suggest that adversarial training is not yet ready for robot learning.
Publishing Year
Date Published
2021-01-01
Proceedings Title
2021 IEEE International Conference on Robotics and Automation
Acknowledgement
M.L. and T.A.H. are supported in part by the Austrian Science Fund (FWF) under grant Z211-N23 (Wittgenstein Award). R.H. and D.R. are supported by Boeing and R.G. by Horizon-2020 ECSEL Project grant no. 783163 (iDev40).
Page
4140-4147
Conference
ICRA: International Conference on Robotics and Automation
Conference Location
Xi'an, China
Conference Date
2021-05-30 – 2021-06-05
ISBN
ISSN
eISSN
IST-REx-ID
Cite this
Lechner M, Hasani R, Grosu R, Rus D, Henzinger TA. Adversarial training is not ready for robot learning. In: 2021 IEEE International Conference on Robotics and Automation. ICRA. ; 2021:4140-4147. doi:10.1109/ICRA48506.2021.9561036
Lechner, M., Hasani, R., Grosu, R., Rus, D., & Henzinger, T. A. (2021). Adversarial training is not ready for robot learning. In 2021 IEEE International Conference on Robotics and Automation (pp. 4140–4147). Xi’an, China. https://doi.org/10.1109/ICRA48506.2021.9561036
Lechner, Mathias, Ramin Hasani, Radu Grosu, Daniela Rus, and Thomas A Henzinger. “Adversarial Training Is Not Ready for Robot Learning.” In 2021 IEEE International Conference on Robotics and Automation, 4140–47. ICRA, 2021. https://doi.org/10.1109/ICRA48506.2021.9561036.
M. Lechner, R. Hasani, R. Grosu, D. Rus, and T. A. Henzinger, “Adversarial training is not ready for robot learning,” in 2021 IEEE International Conference on Robotics and Automation, Xi’an, China, 2021, pp. 4140–4147.
Lechner M, Hasani R, Grosu R, Rus D, Henzinger TA. 2021. Adversarial training is not ready for robot learning. 2021 IEEE International Conference on Robotics and Automation. ICRA: International Conference on Robotics and AutomationICRA, 4140–4147.
Lechner, Mathias, et al. “Adversarial Training Is Not Ready for Robot Learning.” 2021 IEEE International Conference on Robotics and Automation, 2021, pp. 4140–47, doi:10.1109/ICRA48506.2021.9561036.
All files available under the following license(s):
Creative Commons Attribution-NonCommercial-NoDerivs 3.0 Unported (CC BY-NC-ND 3.0):
Link(s) to Main File(s)
Access Level
Open Access
Export
Marked PublicationsOpen Data ISTA Research Explorer
Web of Science
View record in Web of Science®Sources
arXiv 2103.08187