Learning-Based Quadruped Robot Framework for Locomotion on Dynamic Rigid Platforms

Abstract

Typical robot controllers assume firm ground, limiting their effectiveness in controlling robots on dynamic platforms such as trucks or ships. To address this limitation, we propose a reinforcement learning framework for robot locomotion on dynamic rigid platforms and a simulation in which 6-DoF dynamic platforms emulating ship oscillation. The framework enables a reinforcement learning model to estimate platform motion during robot locomotion control. In the simulation, our framework significantly reduces the quadruped robot’s fall rate and trajectory deviation compared to baseline controllers. Experiments on a real robot show that our framework enabled a quadruped robot to adapt to platform motions, including those that threw the robot into the air, while baseline models struggled in this case. Thus, our framework can advance the deployment of robots in real-world marine and vehicular applications.