Legged robots pose one of the greatest challenges in robotics. Dynamic and agile maneuvers animals cannot be imitated by existing methods that are crafted humans. A compelling alternative is reinforcement learning, which requires minimal craftsmanship promotes natural evolution a control policy. However, so far, learning research for legged mainly limited to simulation, only few comparably simple examples have been deployed on real systems. The primary reason training with robots, particularly dynamically balancing systems, complicated expensive. In present work, we introduce method neural network policy simulation transferring it state-of-the-art system, thereby leveraging fast, automated, cost-effective data generation schemes. approach applied ANYmal robot, sophisticated medium-dog-sized quadrupedal system. Using policies trained machine achieves locomotion skills go beyond what had achieved prior methods: capable precisely energy-efficiently following high-level body velocity commands, running faster than before, recovering from falling even complex configurations.

Learning agile and dynamic motor skills for legged robots