Pole starts hanging down - agent learns to swing up and balance. It keeps learning from every step, so change the physics and watch it adapt: lengthen the pole and its return drops, then climbs back within 10-20 episodes. Turn learning off to compare.