State Space

The set of all possible states of a system; for a robot, usually made up of joint angles, pose, velocity, and similar quantities.

In reinforcement learning and control, a state is a set of quantities that describes a system's current situation well enough to predict how it will change next, and the state space is the collection of all possible states. A Markov Decision Process (MDP, the standard mathematical framework for reinforcement learning) writes this as S, which can be discrete, such as squares on a chessboard, or continuous, a vector of real numbers. A robot's state typically includes joint angles and angular velocities, end-effector pose, gripper opening, and the positions of nearby objects. A real robot often cannot access the full state and can only obtain an “observation” through cameras and sensors, in which case the problem becomes a Partially Observable MDP. In VLA papers, “state” usually refers to the robot's own joint values, fed into the model alongside images. Note this is unrelated to “state space models” such as Mamba.

ExampleFor a 7-degree-of-freedom robot arm with a parallel gripper, the proprioceptive state might be 7 joint angles, 7 joint angular velocities, and 1 gripper opening value — a 15-dimensional real vector in total.

Also called

State

Related

Observation · Action Space · Markov Decision Process · Partially Observable Markov Decision Process · Proprioception · State Space Model

Sources

Markov decision process - Wikipedia
State space (computer science) - Wikipedia