Wheel Loader Scooping Controller Using Deep Reinforcement Learning

This article presents a deep reinforcement learning-based controller for an unmanned ground vehicle with custom-built scooping mechanism. The robot's aim is to autonomously perform earth cycles three degrees of freedom: lift, tilt and the velocity. While majority previous studies on automated processes are based data recorded by expert operators, we present method control wheel loader cycle using learning methods without any user-provided demonstrations. controller's approach actorcritic, Deep Deterministic Policy Gradient algorithm which use map online sensor as input continuously update actuator commands. training policy network done solely in simplified simulation environment virtual physics engine, converges average 65% fill factor from full bucket capacity 5 [sec] time. We illustrate performance trained simulations real-world experiments 3 different inclination angles earth. An additional experiment compared our remote manual human control. Overall, exhibited good terms both achieved visually varying scooped weights 4.1 - 7.2[kg], 5.1 7.1[sec] experimental results confirm ability planner required, indicating that can be used excavation purposes.