Models & learning
Privileged information
Privileged information is a signal available during training, supervision, simulation, or data generation but deliberately unavailable to the deployed policy at inference. It may guide a critic, teacher, oracle, or label generator; a valid deployment claim must show that the acting policy uses only observations available on the real system.
Also known as: privileged state, privileged observations, privileged training information
Updated
The information asymmetry
Simulation exposes variables that real robots may not measure directly: exact object poses, contact states, disturbance forces, complete maps, or future success labels. Treating those variables as privileged means they can help training without being assumed available to the deployed actor.
The asymmetric actor–critic study demonstrates one common pattern. Its critic receives full simulator state while the actor receives rendered RGB-D observations. The actor can therefore be transferred to hardware without requiring the critic's complete state input at execution time.
Privileged teachers and oracles
The idea also applies outside reinforcement-learning critics. A teacher can plan from ground-truth state and supply actions to a vision policy. An oracle can use exact geometry to label feasible tasks or generate successful demonstrations. In Splat2Act, a scripted oracle reads object poses, geometry, and robot state to plan simulated pickups, while the deployed policy receives camera images, proprioception, and language.
Privileged information describes the unavailable signal, not one particular learning algorithm. It may enter through labels, rewards, resets, demonstrations, a critic, or a teacher policy.
The leakage boundary
The important question is whether privileged fields influence training only through the stated supervision route. If exact object pose is accidentally included in a deployed observation vector, evaluation no longer measures the camera-based policy that hardware can run. A similar leak occurs when a simulator success flag is available during a rollout but no corresponding real-world detector exists.
Dataset schemas should therefore label each field as deployable, training-only, evaluation-only, or derived. Train and test loaders should enforce that split rather than relying on documentation alone.
What to report
A robotics result using privileged information should identify the privileged variables, which component consumes them, when access ends, and how the deployed observation space is reproduced on hardware. It should also report whether privileged state selected only successful examples, because that filtering can change the training distribution even when the policy never sees the state directly.
Sources
Related terms
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.
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
Policy distillation
Policy distillation transfers behaviour from one or more teacher policies into a student policy by training the student to match teacher outputs on sampled states. In robotics it can consolidate specialist controllers into one deployable policy, but the student remains limited by the states it visits and the quality and coverage of its teachers.
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.