GENERATING A LABELED 3D BOX DATASET FOR MACHINE LEARNING TO LOCALIZE A BOX

Automatically localizing a box of different sizes using machine learning (ML) requires a dataset to train an ML model. Figure 1 shows SICK’s Visionary‑T 3D camera (Visionary-T - Archive | SICK, n.d.) that natively offers only a binary indication of whether an object matches a predefined box size, leaving position, orientation, and dimensional estimation unaddressed. To bridge this gap, a comprehensive workflow was devised in this current study to capture and annotate point‑cloud data for boxes of different sizes and aspect ratios. Three hundred scenes were recorded under varying distances and viewpoints then processed by an automatic labeling pipeline that inferred each box’s six‑degree‑of‑freedom pose and edge lengths. The automated procedure yielded valid annotations for roughly 25 % of the samples, producing an initial, openly available dataset that could be progressively refined through human verification. Once corrected, the dataset enabled the training of ML models capable of simultaneously estimating position, orientation, and size, thereby facilitating accurate grasp planning and collision‑free manipulation for industrial robotic hands. This resource can accelerate research on 3D perception, lower the entry barrier for experimentation, and pave the way for fully data‑driven box‑handling solutions in automated logistics.