AI-Powered Picking, Without the Training Bottleneck

Overview

This document features an interview with Nick Longworth, Manager for Robotic Solutions at SICK, explaining how the company’s dLocate portfolio leverages AI-powered object localization to streamline and accelerate robotic picking automation. Traditional robotic vision often relies on predefined patterns, CAD models, or geometric features, which struggle with flexible, deformable, overlapping, or irregular items. In contrast, SICK’s AI-based object localization employs deep learning neural networks to identify objects by their learned visual characteristics, making it highly effective in complex, variable environments.

A notable challenge in AI-based automation is the need for extensive model training using large annotated image datasets, which can be time-consuming, costly, and inflexible—especially in settings with frequent product changes or diverse SKUs. SICK addresses this with its dLocate portfolio, which includes three AI algorithms integrated in the PLB robot guidance system: Box, SmartPick, and Anything. Box and SmartPick require limited training (as few as 50 to 100 images), facilitated by SICK’s deepAnnotate software, focusing on relatively flat or deformable items respectively. Anything stands out by using a foundation AI model and natural language prompts to identify objects without the need for training data, beneficial for broad retail and grocery products.

The system uses synchronized 2D imaging and 3D point cloud data. The AI model first identifies object candidates in 2D and then maps these to 3D data to validate and determine precise six-degree-of-freedom (6-DOF) pick coordinates. Beyond localization, PLB offers advanced features like sorting, empty-container detection, fill-level monitoring, collision avoidance, and pickability analysis.

Choosing between the algorithms depends on application needs: Box for flat items, SmartPick for flexible or overlapping items, and Anything for diverse consumer goods. Unlike traditional vision systems with rigid rule-based methods, dLocate’s AI provides flexibility for challenging items that were previously difficult to handle in automation.

If AI detection falls short, users can tweak imaging or revert to rule-based algorithms—PLB supports hybrid approaches to maximize performance and ROI. Overall, SICK’s AI-powered dLocate solutions enable manufacturers to reduce manual annotation and training, lower engineering effort, speed deployment, and open new possibilities for robotic picking in both industrial and retail environments.

For more information, visit www.sick.com.