Models & learning
Observation space
An observation space is the set and representation of measurements available to an agent or robot policy at each decision step. It can contain images, joint state, force or tactile signals, language, estimated poses, or histories of those values, and it need not reveal the complete underlying state of the environment.
Also known as: robot observation space, observation spaces
Updated
Observation is not the same as state
The state contains the information needed to describe the environment for the model being used. An observation is what the agent receives. A humanoid may observe camera images and joint encoders without directly observing object mass, floor friction or another person's intent.
Sutton and Barto distinguish fully observed Markov decision processes from partially observed settings. Stacking recent observations can expose motion or contact history, but it does not guarantee that every hidden variable becomes recoverable.
The space defines an interface
Gymnasium spaces describe the shape, data type and valid values of observations and actions. Robot observations are often structured: an image tensor, a vector of joint values and a task instruction may arrive together while retaining different meanings.
Document field order, units, bounds, coordinate frames, sampling times, normalisation and missing values. A tensor shape alone cannot say whether a position is absolute or relative, whether angles use radians, or whether an image has already been cropped.
Deployment must reproduce the training inputs
A policy trained with simulator-only object poses or perfect contact labels has a different observation space from one deployed with cameras and noisy estimates. Removing privileged inputs after training is safe only when the method explicitly accounts for that change.
Dataset evaluation should separate raw measurements from derived estimates and identify any history window or preprocessing. An observation space is part of the policy contract: changing its ordering, scale, delay or available fields changes the problem even when the robot and task look identical.
Sources
Related terms
Data & collection
Robot state
Robot state is the set of variables used to describe a robot at a particular time, such as joint positions and velocities, base pose, end-effector pose, gripper state, actuator measurements, or estimated motion. In control theory, a complete state contains enough information to predict future evolution given an action; in robot datasets, “state” often means only the measured or estimated subset that was logged.
Models & learning
Action space
An action space is the set and representation of commands that an agent or robot policy is allowed to choose. In robotics, actions may be discrete choices or continuous values such as joint targets, motor torques, end-effector pose changes, base velocities, or gripper commands. The action space defines what the policy outputs, not necessarily the motion the hardware ultimately executes.
Models & learning
Policy
A policy is the decision rule that maps a robot’s current observations or estimated state, and sometimes a task instruction, to an action or probability distribution over actions. It can be hand-designed or learned from demonstrations, rewards or both. In humanoid robotics, its outputs may be joint targets, torques, end-effector changes or higher-level skills.
Hardware & control
Proprioception
Proprioception is sensing of a robot’s own internal configuration and motion rather than the external scene. For a humanoid it commonly includes joint positions and velocities, actuator effort or torque, and inertial measurements of body rotation and acceleration. These signals support state estimation and feedback control but do not, by themselves, directly describe nearby objects or terrain.
Data & collection
Data modality
A data modality is a distinct kind or channel of information characterised by how it is sensed, represented, or interpreted, such as RGB images, depth, language, joint state, actions, audio, tactile measurements, or force–torque signals. A multimodal robot dataset contains more than one modality, but usefulness depends on their alignment, semantics, and relevance to the task.
Hardware & control
Observability
Observability is the property that a system's internal state can be determined from its measured outputs over time, given the known inputs and the assumed model. If different states can generate indistinguishable observations, those differences are unobservable under that sensing setup.