Reinforcement learning

Reinforcement learning (RL) is a machine-learning paradigm where an agent learns by maximizing a reward signal through trial-and-error interaction with an environment. Distinct from imitation learning (learning from demonstrations) and from supervised learning (learning from labeled examples). The defining property: the agent must explore actions and observe consequences before it can learn what works.

The distinction matters because RL in the real world is data-hungry to the point of impracticality for most robot tasks — a humanoid trying to learn manipulation by trial-and-error would damage itself before learning anything useful. Most RL in robotics happens in simulation, with sim-to-real transfer for deployment. Boston Dynamics' Atlas uses RL for locomotion training; Figure's manipulation policies use a combination of RL and imitation. Pure-RL approaches haven't crossed into deployed humanoid programs at scale yet.