Privileged signals (true velocity, foot force, foot height) go to the critic only

Symptom

Policies trained on ground-truth base linear velocity work in sim and fail on hardware, where only a drifting IMU and encoders exist - the policy has learned to depend on a signal that does not survive deployment.

Context

Many open-source locomotion stacks feed simulator ground-truth linear velocity to the actor. The reference team refused: the real robot has no ground-truth velocity. Asymmetric actor-critic keeps the training benefit of privileged information without deploying the dependency.

Change

Route ground-truth velocity, foot contact forces, and foot heights to the critic only; the actor observes exclusively signals that exist on hardware (IMU-derived quantities, encoders, commands, previous actions).

Outcome

Recorded as adopted doctrine in Lucen's experience log; the trained actor's input contract matches what the real robot can actually produce.

Mechanism

The critic is discarded at deployment, so it may consume any privileged state to reduce value-estimation variance; the actor's observation set is a deployment contract - anything in it that hardware cannot supply (or supplies with different noise/drift) becomes a train/deploy distribution shift the policy was never trained to handle.

Applies when

• designing actor/critic observation spaces
• reviewing a config where the actor sees base_lin_vel or contact forces
• sim policy is strong but real robot drifts, oscillates, or falls without obvious actuator cause