Learning Humanoid Balance on an Unstable Support: Effects of Action Representation and Sensory Constraints

A robot balancing on an unstable support has only the local, noisy signals biology works from. We ask what that restriction costs a learned controller. Three pipelines—PD-assisted inverse kinematics (IK), IK alone and direct joint-space control—were trained with Proximal Policy Optimization on a humanoid on a wobble board, under privileged simulator state and under onboard-measurable observations, with the seed as the replicate. IK converts a desired foot pose into the joint angles that realize it, so two pipelines command a posture and the third commands joints directly. With privileged state and the distribution mean, PD-assisted IK completed 78% of 60 s trials against 35% and 34%; restricted to onboard sensing it completed none while the other two reached 82% and 62%; sampling actions reverses the privileged ordering. At the mildest noise tested every restricted-observation policy completed no trial, while two of the three privileged configurations were unaffected or better; one control step of whole-input delay left almost no completed trial anywhere. We therefore recommend no single pipeline: evaluate one under the observations and action-selection rule it will run with, and report nominal success alongside tolerance to noise, latency and plant error.