Automating Box Folding: Sequence Extraction and Ranking Methodologies

Technical Summary: Automating Box Folding: Sequence Extraction and Ranking Methodologies

Problem Statement
The paper addresses the challenge of automating the box folding process within industrial packaging systems. While significant research exists on optimizing space filling and package stability, the specific task of determining how to fold a carton remains under-explored. Existing literature typically bifurcates into two extremes: theoretical approaches that derive all possible folding sequences without regard for hardware constraints, and hardware-specific solutions tailored to particular carton types that lack flexibility. This work seeks to bridge this gap by providing a systematic framework that identifies feasible folding sequences for any carton design while simultaneously evaluating their implementability on specific robotic hardware.

Methodology
The proposed approach is a three-phase methodology designed to transition a flat carton from an initial state ($S_0$) to a final configuration ($S_f$):

1. Box Modeling and Connectivity Matrix Generation:
   The carton is modeled as a kinematic mechanism where panels act as rigid links and folds act as revolute joints. The system utilizes a connectivity matrix ($C$) to encode adjacency and hierarchical dependencies (parent-child relationships) between panels. The state of the carton is defined by the joint angle vector ($\theta$) and the connectivity matrix, represented dynamically as a connected graph ($G_t$). Physical properties, including panel dimensions and geometric centers, are incorporated to enable precise spatial calculations.

2. Folding Sequence Identification:
   The folding process is treated as a sequence of discrete actions where, at each step, only one joint is rotated from its initial to its final angle. The system employs a decision tree representation explored via a backtracking algorithm (depth-first search). A collision-checking function ($CC$) validates each incremental rotation. A sequence is recorded only if it transitions the carton from $S_0$ to $S_f$ without violating geometric constraints or causing collisions. This phase generates a set of all theoretically feasible sequences ($\sigma_1, \dots, \sigma_q$) independent of specific hardware.

3. Evaluation Metrics for Ranking:
   To bridge the gap between theoretical feasibility and practical implementation, the authors introduce specific cost functions to rank the identified sequences based on hardware compatibility:

   • Bounding Box Volume Criterion ($C_{vol}$): Minimizes the cumulative spatial volume occupied by the carton throughout the folding process to promote compactness.
   • Bounding Box Maximum Dimension Criterion ($C_{dim}$): Minimizes the largest single dimension of the bounding box at any step, ensuring compatibility with robotic workspace limitations.
   • Number of Aerial Foldings Criterion ($C_{aerial}$): Counts folds performed without workbench support ("Aerial Folds"). Minimizing this metric enhances stability and repeatability, particularly in setups lacking external vision feedback.

Key Contributions

• Generalized Sequence Extraction: The paper proposes an algorithm that generates all feasible folding sequences for any carton design without being pre-tailored to specific hardware, moving beyond the limitations of fixture-based or purely theoretical models.
• Hardware-Aware Ranking: A novel contribution is the introduction of evaluation metrics that rank sequences based on the constraints of the available automation hardware (e.g., robot reach, stability requirements).
• Systematic Framework: The work provides a complete pipeline from kinematic modeling to the selection of a specific, implementable folding path.

Results and Case Study
The methodology was validated through a case study using an ABB GoFa CRB 15000 collaborative robot equipped with a custom tool featuring suction cups and spring mechanisms. The setup included a workbench with aluminum profiles and L-shaped supports to stabilize panels during folding.

• Simulation: The system was implemented in Unity to handle complex 3D collision detection and kinematic modeling.
• Sequence Selection: The algorithm generated over 100 valid sequences. Using the proposed metrics, sequences were ranked primarily by minimizing aerial folds and secondarily by minimizing maximum box dimensions.
• Validation: The top-ranked sequences were tested against the physical constraints of the robot. The first five sequences were deemed unfeasible because their starting fold points exceeded the robot's workspace. The sixth sequence, however, satisfied all constraints and was successfully executed, transforming a flat box into a stable structure ready for object insertion.

Significance and Claims
The authors claim that this work offers a flexible and efficient solution that addresses a specific gap in the literature: the lack of a method that both identifies all possible folding paths and evaluates them against real-world hardware constraints. By generalizing the process for any carton design and introducing metrics to prioritize stability and workspace compatibility, the research provides a systematic framework for industrial implementation. The paper concludes modestly, noting that while the case study demonstrates feasibility, further statistical validation across multiple box types is required to comprehensively prove that the method consistently identifies feasible solutions in fewer attempts.