PAC-Bayes Regularization for Learning Controllers that Generalize Across Environments

Abstract

We bound the error rate of a robotic grasping controller in novel environments by connecting recent work involving PAC-Bayes regularization to algorithmic decision making. Training a controller on the KUKA robot in the Bullet physics simulator, we compute a PAC-Bayes on the learned controller and show generalization bounds on the performance of the controller in new environments.