Action Space

The set of all actions an agent can take at each step, and the numerical form those actions take.

Action space is a foundational concept in reinforcement learning and robot learning: the full set of actions an agent can choose from at each step. There are two kinds. A discrete action space has a finite number of options, like the handful of buttons in an Atari game. A continuous action space consists of real-valued numbers, such as the target angles for a robot arm's joints, or the displacement and gripper opening of the end effector (the gripper or tool at the very tip of the arm). Robot control at the lowest level is almost always continuous, though discrete examples exist too — embodied-navigation benchmarks often use just a few actions such as move forward, turn left, turn right, and stop. The same task can be represented with joint angles or end-effector pose, and with absolute or incremental (delta) values; different robots also have different action dimensions, which is exactly the mismatch cross-embodiment training has to handle. Vision-language-action (VLA) models commonly discretize continuous actions into tokens for output, or generate continuous values directly via diffusion, flow matching, or straightforward regression.

ExampleThe CartPole task has just two actions, push left and push right — a discrete action space. OpenVLA outputs robot-arm actions as 3D translation plus 3D rotation plus 1D gripper open/close, a 7-dimensional continuous action space.

Also called

Continuous Action Space, Discrete Action Space

Related

State Space · Observation · Policy · Delta (Relative) Action vs. Absolute Action · Action Tokenizer · Unified Action Space

Sources

OpenAI Spinning Up: Key Concepts in RL
Gymnasium Documentation: Basic Usage