Gait-parameterized RL for locomotion across morphologies and scales without reward retuning

Abstract

Reinforcement learning (RL) has enabled agile legged locomotion; however, transferring these results to a new robot often requires extensive reward redesign and coefficient retuning as the morphology, scale, and actuation constraints change. This study evaluates a gait-parameterized RL formulation for cross-platform locomotion learning without reward retuning. We keep the reward terms and coefficients, network architecture, and PPO hyperparameters fixed and adjust only the platform interface via observation/action mappings and low-level actuation settings. With this setup, we trained separate policies in the simulation on three small quadrupeds (Go1, ANYmal C, and Mini Cheetah), a humanoid (Atlas), and a large-scale hydraulic quadruped (BeTheX-Q), and evaluated their performance across multiple terrains. The results showed stable locomotion across large variations in morphology and physical scale. The learned gait-intent outputs further exhibited systematic scaling and correlation trends with speed, posture, actuation effort, and mechanical cost of transport, thereby providing an interpretable view of embodiment-dependent behaviors under a fixed learning formulation. Finally, we extend the framework to a wheeled-legged hybrid configuration of BeTheX-Q and demonstrate wheel–leg coupled locomotion in both simulation and hardware.