Reinforcement learning is a way to train behavior by rewarding the outcomes you want. A robot or simulated agent tries actions, receives feedback, and gradually learns a policy that produces better results.
In robotics, reinforcement learning is often useful for movement: balance, walking, recovery, grasp attempts, or control policies that are hard to hand-code. It can produce surprisingly robust behavior, especially when paired with simulation.
The catch is that reward design is difficult. If the reward is too simple, the system may learn a shortcut that looks good in training and fails in reality. If the training environment is too narrow, the learned behavior may not generalize.
When a humanoid company mentions reinforcement learning, treat it as a method, not a magic explanation. Ask what behavior was trained, whether it transferred to hardware, how much human demonstration data was used, and whether the result works outside a controlled test.