Deep Reinforcement Learning-Based Obstacle Avoidance for Embedded Mobile Robot Navigation

Abstract

This paper presents the design, implementation, and deployment of an autonomous robot navigation system using a Deep Q-Network (DQN) reinforcement learning algorithm on the NVIDIA Jetson Nano platform. This work examines the integration of Edge AI techniques for real-time decision-making in navigation scenarios with limited sensory input. The robot employs three ultrasonic sensors for environmental perception, combined with a lightweight DQN model optimized for embedded deployment. The system incorporates TensorRT optimization, model quantization, and hardware-aware task distribution across the CPU and GPU. Experimental results show a collision rate of 4.8 percent, an inference latency of 17.1 ms, and a power consumption below 10 W, with a model size of 493 KB.