Performance Analysis and Optimization of Hybrid Edge-Cloud Architectures for Real-Time Robotics Applications

Abstract

This thesis explores the optimization of distributed robotic perception systems for real-time applications such as autonomous navigation, smart surveillance, and multi-agent coordi- nation. These systems require fast processing of high-frequency sensor data under tight latency and reliability constraints. Edge computing offers low-latency inference and privacy, while cloud computing pro- vides centralized scalability and resource efficiency, but suffers from transmission delays and network dependency. Critically, the bandwidth required to transmit raw video streams and LiDAR pointcloud messages to the cloud is often infeasible, necessitating edge-side preprocessing. To address these challenges, this work proposes and evaluates a hybrid edge–cloud architecture in which each sensor node that is equipped with its own LiDAR and camera performs background removal and motion detection locally at the edge, while the remaining perception tasks are offloaded to the cloud. This design reduces bandwidth usage and enables real‑time responsiveness under constrained conditions. While edge modules cannot perform full object classification, they provide fast ”reflex-like” responses that enable event filtering, alert triggering, and resource prioritization, reserving the more computationally intensive object recognition for cloud processing. LiDAR processing was parallelized using Intel TBB and spatial chunking, achieving up to 2× speedups across 1–32 cores. For camera-based perception, Frame Differencing, GMM, and Dense Optical Flow were tested. Frame Differencing proved most effective for edge deployment, achieving 100% message reliability and 80.9% bandwidth savings with a 20.8 ms average processing time. Scalability tests showed that a 64-core system can support up to 75 nodes at 10 Hz or 50 nodes at 20 Hz with no message loss while meeting real-time constraints. Communication protocol testing revealed latencies ranging from 0.24 ms (ROS) to 2,589 ms (ZeroMQ over WiFi), setting architectural limits for cloud offloading. Finally, cost analysis showed that a 64-core cloud instance could replace 50 edge devices (￿$44,950 upfront), offering cost- effective scalability in suitable deployments. This work delivers: (1) empirical benchmarks that reveal how LiDAR and camera perception scale under parallel processing, (2) motion‑based filtering techniques that sig- nificantly reduce bandwidth without sacrificing accuracy, (3) real‑world measurements of communication protocol latency under 5G and Wi‑Fi, and (4) practical deployment guide- lines for hybrid edge–cloud robotic systems.