Quantifying generalization of RL algorithms on DeepMind Control Suite

Algorithms trained on a single environment with no visual randomization perform poorly on new test sets with variations like randomizing object colors or textures./nThe problem of learning a policy that can solve the same task in all environments becomes much more difficult with domain randomization./nTraining algorithms on all existing environments is impractical and challenging./nData augmentation and randomization alone are not scalable solutions for generalization./nGaussian noise as a form of data augmentation does not generalize to different types of noise./nDefining all necessary data augmentation techniques ahead of time is challenging and may not truly facilitate learning in machine learning algorithms.