Robotic pick-and-place applications require vision systems that can convert 3D sensor data into reliable object poses while meeting strict constraints on latency, installation space, and processing resources. This paper presents a sensor-agnostic embedded 3D vision architecture for AI-assisted object localization, robot guidance, and additional 3D inspection tasks, such as weld inspection.

The system accepts depth information from different 3D cameras and sensors and converts it into a common point-cloud representation. AI-based methods identify target objects and distinguish them from their surroundings, while deterministic 3D algorithms calculate their position and orientation. The resulting pose is transformed from the camera coordinate system into the robot coordinate system and transmitted to the robot for pick-and-place operations or inspection positioning.

The paper examines the complete processing chain, including 3D data acquisition, point-cloud reduction, object recognition, pose estimation, calibration, coordinate transformation, and robot communication, with all processing executed locally on the edge device. Particular attention is given to the distribution of tasks between neural inference and geometric 3D processing under limited computing resources. The same workflow can also integrate additional identification tasks, such as code reading, before or after object handling.

The architecture is implemented on the ORBIS embedded 3D vision platform. A robotic pick-and-place demonstrator and a weld-inspection use case illustrate how local processing, sensor-independent acquisition, and combined AI and deterministic vision methods can support adaptable robot guidance and inspection without relying on an external PC.

This version keeps robot guidance and AI-based localization as the main subject, while weld inspection and code reading demonstrate that the embedded architecture can support additional inspection and identification tasks.