Deterministic edge intelligence for low-latency robotic control

Abstract

This paper presents a distributed Edge AI architecture for the real-time, contactless teleoperation of a 6-DOF robotic manipulator using natural hand gestures. In environments with strict sterility or safety requirements, such as cleanrooms or hazardous industrial zones, traditional physical interfaces are impractical. The proposed system addresses this challenge without relying on costly workstations equipped with discrete GPUs or cloud computing. The solution leverages a budget-friendly Raspberry Pi 4 paired with an OV9281 global shutter camera, effectively eliminating motion blur during rapid hand movements at up to 30 frames per second. Local video stream processing is performed using the lightweight MediaPipe Hands framework deployed entirely on the CPU. To ensure high reliability and avoid the misclassifications common in ultra-lightweight neural network classifiers, gesture classification is executed through a deterministic geometric analysis of 21 extracted hand landmarks, reducing inference time to under 0.2 milliseconds. The extracted control commands are integrated into a service-oriented ecosystem and transmitted to a ROS 2 Humble environment through an optimized multi-layer pipeline: a UART interface, a Raspberry Pi Pico microcontroller bridge, a custom WebSocket server, and a USB/IP network forwarder into a WSL2 environment. Experimental results demonstrate an end-to- end latency of under 40 ms. This confirms that the engineered architecture provides stable, low-latency control of the physical manipulator while maintaining complete data privacy and system resilience.