One-Shot Box-Centric Teaching for Persistent Robotic Sorting-and-Filling with Relative Pose Constraints

Area of Science:

• Robotics
• Computer Vision
• Manufacturing Automation

Background:

• Flexible manufacturing demands robots adapt to variations in container poses, object availability, and sensing conditions for sorting and filling tasks.
• Existing methods often require extensive training or struggle with dynamic environmental changes.

Purpose of the Study:

• To propose a box-centric one-shot teaching framework for robotic packing tasks that preserves relative pose constraints.
• To enable robots to reproduce specified in-box arrangements with minimal human intervention.

Main Methods:

• A one-shot teaching stage utilizes reference-prompted SAM-based contour refinement to extract object details and relative pose constraints.
• Constraints are converted into a reusable box-centric packing template, preserving translational and angular layout information.
• A perception-to-command pipeline generates pick-and-place commands, with continuous assignment and state updates for persistent execution.

Main Results:

• Single-box template transfer experiments achieved low mean placement errors (7.16 mm and 7.57 mm), visually preserving relative object orientations.
• Multi-box experiments demonstrated successful completion of residual slots after scene updates without re-teaching.
• Validation across different container types and object shapes confirmed framework extensibility; SAM refinement enhanced template acquisition robustness.

Conclusions:

• The proposed framework enables one-shot template acquisition and robust box-centric layout transfer for constrained robotic packing.
• It facilitates persistent task-level execution by preserving relative pose constraints and adapting to dynamic environments.
• This approach significantly enhances the adaptability and efficiency of robotic sorting and filling in flexible manufacturing.