Sensor-Based Navigation Using Hierarchical Reinforcement Learning

Robotic systems are nowadays capable of solving complex navigation tasks. However, their capabilities limited to the knowledge designer and consequently lack generalizability initially unconsidered situations. This makes deep reinforcement learning (DRL) especially interesting, as these algorithms promise a self-learning system only relying on feedback from environment. In this paper, we consider problem lidar-based robot in continuous action space using DRL without providing any goal-oriented or global information. By solely local sensor data solve tasks, design an agent that assigns its own waypoints based intrinsic motivation. Our is able learn goal-directed behavior even when facing sparse feedback, i.e., delayed rewards reaching target. To address challenge complexity space, deploy hierarchical structure which exploration distributed across multiple layers. Within structure, our self-assigns internal goals learns extract reasonable reach desired target position data. experiments, demonstrate two environments show seriously improves performance terms success rate weighted by path length comparison flat structure. Furthermore, provide real-robot experiment illustrate trained can be easily transferred real-world scenario.