Domain Randomization
Domain randomization (DR) is a technique used to improve the robustness and generalizability of machine learning models, particularly in robotics and computer vision, by training them on simulated environments with artificially introduced variations in parameters like lighting, textures, and object properties. Current research focuses on optimizing DR strategies, including developing methods for efficient parameter selection and leveraging techniques like Bayesian optimization and continual learning to improve sample efficiency and avoid overly conservative policies. The impact of DR is significant, enabling more reliable transfer of models from simulation to real-world applications, reducing the need for extensive real-world data collection and improving the performance of systems operating in unpredictable environments.
Papers
Safe Continual Domain Adaptation after Sim2Real Transfer of Reinforcement Learning Policies in Robotics
Josip Josifovski, Shangding Gu, Mohammadhossein Malmir, Haoliang Huang, Sayantan Auddy, Nicolás Navarro-Guerrero, Costas Spanos, Alois KnollTechnical University of Munich●Berkeley●Technische Universit¨at Berlin●Leibniz Universit ¨at HannoverLearning Robotic Policy with Imagined Transition: Mitigating the Trade-off between Robustness and Optimality
Wei Xiao, Shangke Lyu, Zhefei Gong, Renjie Wang, Donglin WangWestlake University