Sim-to-Real for Robot Reinforcement Learning
Robotics at Google is a research team devoted to exploring how machine learning can revolutionize the world of robotics. We are interested if we can train policies in simulation and transfer them to real robots. This transfer involves dealing with the differences of idealized simulations and the complexities of the real world, also known as the simulation reality gap.
We discuss our research and experiences in dealing with our sim-to-real research applied to robotic grasping and locomotion control of legged robots. In both cases, we need a deep understanding of dynamics and environment through sensing and learning and use this knowledge to build a simulation that is sufficiently rich to allow the transfer of policies. We also discuss various approaches to use machine learning to automatically deal with the reality gap.Registration