ROS-Based Visual Perception Using an M.2 Edge AI Accelerator for Embedded Robotics

Embedded robotic systems increasingly rely on deep neural networks for visual perception tasks such as object detection and scene understanding. These models often exceed the computational capabilities of conventional embedded processors. Edge AI accelerators address this challenge by providing specialized hardware for efficient neural network inference. However, many accelerator software stacks are designed for high-throughput streaming pipelines, while robotics middleware such as the Robot Operating System (ROS) typically operates using frame-triggered, message-driven execution. This thesis investigates the system-level integration of the Axelera Metis M.2 edge AI accelerator into a ROS-based perception pipeline running on an embedded NVIDIA Jetson Orin platform. A frame-triggered integration strategy is proposed that enables synchronous accelerator inference within ROS callbacks without modifying vendor runtime internals, preserving ROS event-driven execution while maintaining compatibility with the accelerator’s stream-oriented software stack. Accelerator-backed inference nodes were implemented in both Python and C++, and the perception pipeline was evaluated through latency measurements covering preprocessing, accelerator inference, and postprocessing stages. Results show stable and predictable inference execution, while overall latency is dominated by host-side processing rather than accelerator runtime. The approach was validated in a robotic manipulation scenario where detections guided pick-and-place actions. Results demonstrate reliable perception outputs for downstream motion planning. Overall, the work shows that stream-oriented AI accelerators can be effectively integrated into frame-triggered robotics middleware and should be evaluated within complete perception pipelines rather than isolated inference benchmarks.