Edge AI to Edge Robotics: Enhancing Human Pose Estimation with High-Performance TPU Computing

Edge AI to Edge Robotics: Enhancing Human Pose Estimation with High-Performance TPU Computing

연구 분야: Artificial Intelligence

학회: International Conference on Advanced Network Technologies and Intelligent Computing

The convergence of Edge AI and high-performance computing (HPC) is revolutionising intelligent systems by enabling real-time data processing and decision-making at the network’s edge. The demand for high-performance computing in resource-constrained environments opens opportunities for applications such as autonomous vehicles, smart cities, and industrial automation. This study presents the terminology that evidences the evolution of edge AI to the edge Robotics landscape, and then we present the novel approach to leveraging Tensor Processing Units (TPUs) to enhance the inference capabilities of robotic systems at the edge. Using a Reachy robot as a case study, we demonstrate real-time human pose estimation powered by TPUs. The research investigates the intersection of robotics, Edge AI, and HPC, focusing on accelerating inference tasks with TPUs. We implemented deep learning-based human pose estimation models—PoseNet-ResNet50, PoseNet-MobileNet V1, MoveNet Lightning, and MoveNet Thunder—on the Robot Operating System (ROS) and benchmarked their performance. Results show that MoveNet Thunder achieved the fastest inference time (48 ms) with a pose score of 98%, while PoseNet-ResNet50 was the slowest (701 ms) with a score of 93%. Our findings highlight the significant improvements in inference speed and accuracy using TPUs over CPUs. We discuss the implications of TPU-accelerated AI in robotics and explore future research directions in this evolving field.