Domain Randomization Sensitivity Analysis for Sim-to-Real Transfer of RL-based Locomotion on Low-Cost Actuator Platforms: A Framework and Experimental Protocol

Domain randomization (DR) is the standard remedy for the sim-to-real gap in reinforcement-learned locomotion, yet existing DR sensitivity studies are confined to high-precision platforms whose near-ideal actuators sideline actuator error. On low-cost hardware, actuator and contact modeling error plausibly dominate the transfer gap, but no systematic method characterizes this regime. We present a framework and experimental protocol for DR sensitivity analysis targeting low-cost actuator platforms, comprising: (i) a one-at-a-time ablation protocol that ranks randomized parameters by their marginal effect on transferred gait performance; (ii) an extension of legged_gym adding actuator-parameter randomization (PD gains and torque limit), absent from the public codebase; and (iii) three evaluation metrics with a hardware validation protocol. We instantiate the framework on a six-joint MI-motor biped with a JY901 IMU, specify randomization ranges from measured hardware parameters, and hypothesize that non-rigid-body parameters (contact friction, PD stiffness) dominate rigid-body parameters (CoM offset, inertia, mass). The protocol designed to test this hypothesis is detailed.