AGPNet - Autonomous Grading Policy Network

In this work, we establish heuristics and learningstrategies for autonomous control of a dozer grading an unevenarea studded with sand piles. We formalize the problem as aMarkov Decision Process, design a simulation which mimicsdozer-soil interactions and finally, compare our simulator to areal-world scaled environment. We use methods from reinforcementlearning, behavior cloning and contrastive learning to traina hybrid policy. Our trained agent, AGPNet, reaches human-levelperformance and outperforms current state-of-the-art machinelearning methods for the autonomous grading task. In addition,we show that our agent is capable of generalizing from randomscenarios to unseen real world problems.