A quadruped robot has learned to walk across slippery, uneven terrain entirely through simulation, without any human-designed gaits or manual tuning. The system relies on deep reinforcement learning ...