humanoidsdata.com

Search

Search datasets, articles, and glossary terms for humanoids and embodied AI.

← All glossary terms

Models & learning

Reinforcement learning

Reinforcement learning is a method in which an agent learns a policy by interacting with an environment and optimising cumulative reward. In humanoid robotics, actions change the robot and world, while observations, rewards and episode endings provide experience for improving balance, locomotion or manipulation behaviour.

Also known as: RL

Updated

Learning from interaction

An agent observes its situation, selects an action through a policy, receives a new observation and reward, then repeats. The objective is expected cumulative reward rather than accuracy against a fixed action label. Sutton and Barto describe trial-and-error search and delayed reward as defining features: an action may matter because of consequences many steps later.

The resulting training data consists of trajectories containing observations, actions, rewards and boundary signals. Some methods learn while gathering fresh experience; others reuse stored experience. Policy-gradient algorithms update a parameterised policy from sampled trajectories. Proximal Policy Optimization is one widely used formulation that limits how far an update moves the policy.

What it means for humanoids

Humanoid control has large action spaces, coupled joints and strict contact constraints. Reinforcement learning can optimise behaviours that are difficult to specify as individual target actions, including recovery, locomotion and transitions between skills. Reward design remains part of the task specification: an incomplete reward can favour motion that scores well without being safe or useful.

Many physical trials are costly and risky, so training commonly begins in simulation. DeepMimic showed how reinforcement learning could train simulated humanoid characters to track motion-capture clips while satisfying task objectives. A policy intended for hardware still has to tolerate differences in dynamics, sensing and latency. Demonstrations, motion priors and conventional safety controllers can constrain exploration, but they do not remove the need for real-robot validation.

Sources