DeepMind reports sim-to-real transfer without fine-tuning

Policies trained entirely in simulation held up on hardware in 61 percent of trials.