+ All Categories
Transcript
Page 1: AdversariallyRobust Policy Learning through Active ...web.stanford.edu/.../2017...Policy_Learning_Poster.pdf · During policy evaluation, we collect N trajectories via policy rollouts.

Adversarially RobustPolicyLearningthroughActiveConstructionofPhysically-PlausiblePerturbations

AjayMandlekar,Yuke Zhu,Animesh Garg,LiFei-Fei,SilvioSavareseDepartmentofComputerScience,StanfordUniversity

Introduction

Demonstrated Robustness in Physical Dynamics Parameters

ARPL Algorithm

Physically-Plausible Threat Model

Experimental Setup

References

ARPL Agent Examples

Top Related