Real-Time Implementation and Evaluation of Object Detection and Obstacle Avoidance Systems for UAVs on NVIDIA Jetson Nano and Raspberry Pi 4 with Coral TPU

Abstract

Current developments in embedded machine learning have unveiled new possibilities in autonomous navigation, especially in the context of unmanned aerial systems (UAVs). With UAVs depending more on onboard systems for real-time perception and reasoning, hardware platform selection is now of utmost importance. This study performs an in-depth benchmarking and performance analysis of embedded edge-computing platforms, i.e., the NVIDIA Jetson Nano, and the Raspberry Pi 4 with Coral TPU, for the deployment of real-time object detection and obstacle avoidance systems in unmanned aerial vehicle (UAV) applications. The central system architecture applies YOLOv11n visual object detection with a ViT-based Proximal Policy Optimization (PPO) reinforcement learning agent to enable semantically driven decision making. All the testing was done in a simulation of the UAV operating in a CoppeliaSim reconstructed cluttered environment. The work emphasizes key performance metrics such as inference time, energy consumption, reward consistency, and collision rate. These indicators were used to measure real-time performance, computational speed, and the stability of learned policies in obstacle-rich spaces. Experimental results showed that the Jetson Nano, with TensorRT optimization and GPU, consistently reported better performance with lower inference latency of 14.03s and 17% better than the Raspberry Pi with TPU, in terms of collision avoidance, but at the cost of higher energy expenditure of 167J. Conversely, the Raspberry Pi 4 with the Coral TPU provided enhanced energy efficiency but it was outperformed because of its lower inference speed and control action instability with frequent collisions. Overall, this study introduces a novel implementation strategy where YOLOv11n object detection outputs that are contextually enhanced using ViT embeddings prior to input to the PPO in contrast to providing raw image inputs. Additionally, all the models were custom trained and their deployment was validated in CoppeliaSim, integrated via a ZMQ-based remote API. The comparative analysis is the first to benchmark an end-to-end detection and avoidance algorithms for UAV system using real embedded hardware. It provides useful advice to UAV makers, roboticists, and embedded system designers in identification of the edged AI computing devices according to their performance strength.