Curriculum learning for safe mapless navigation

Abstract

This work investigates the effects of Curriculum Learning (CL)-based\napproaches on the agent's performance. In particular, we focus on the safety\naspect of robotic mapless navigation, comparing over a standard end-to-end\n(E2E) training strategy. To this end, we present a CL approach that leverages\nTransfer of Learning (ToL) and fine-tuning in a Unity-based simulation with the\nRobotnik Kairos as a robotic agent. For a fair comparison, our evaluation\nconsiders an equal computational demand for every learning approach (i.e., the\nsame number of interactions and difficulty of the environments) and confirms\nthat our CL-based method that uses ToL outperforms the E2E methodology. In\nparticular, we improve the average success rate and the safety of the trained\npolicy, resulting in 10% fewer collisions in unseen testing scenarios. To\nfurther confirm these results, we employ a formal verification tool to quantify\nthe number of correct behaviors of Reinforcement Learning policies over desired\nspecifications.\n