"Scaling Sim2Real Reinforcement Learning" – Osbert Bastani, TTIC Colloquium

“Scaling Sim2Real Reinforcement Learning” Osbert Bastani, University of Pennsylvania Originally recorded on September 28, 2026, at TTIC. In this talk, Osbert Bastani explores how simulation-to-real-world (Sim2Real) reinforcement learning can advance the development of generalist robotics. He discusses two major challenges to scaling Sim2Real RL: reducing the manual effort required to design simulation environments and developing reinforcement learning algorithms specialized for real-world transfer. The talk highlights the use of large language models to generate simulation environments, reward functions, training curricula, and articulated assets, alongside new approaches for training flow policies with the goal of distilling them into vision-language-action (VLA) foundation models. #ReinforcementLearning #Sim2Real #Robotics #MachineLearning #ArtificialIntelligence #LLMs #VisionLanguageAction #FoundationModels #DeepLearning #TTIC