Generalizable dual-arm fabric folding via parallel multiresolution attention

Purpose This study aims to address the challenges of robotic cloth folding, stemming from complex dynamics and high degrees of freedom. While existing learning-based approaches have shown promise, they often suffer from limited generalization across diverse fabrics and require extensive real-world data. To address this gap, we propose a perception-centric strategy introducing HrcbamFolding, a dual-arm system that leverages a deep network to directly map visual inputs to the key manipulation points required. This approach simplifies the complex problems of state estimation and motion planning into a structured keypoint detection task, effectively bypassing the need for explicit physical modeling of the fabric. Design/methodology/approach This study proposes HrcbamFolding, which combines a multiresolution neural network with a channel–spatial attention mechanism to spotlight task-critical fabric regions, thereby enhancing visual-perception accuracy and generalization across diverse materials. It further uses a grasp-pose prediction module that translates visual inputs directly into coordinated grasping and placement actions for each arm, which reduces motion-planning errors and improves execution efficiency. The framework is trained purely in simulation on rectangular fabrics before being assessed on three multistep folding benchmarks. Findings This study shows that in all three tasks, HrcbamFolding achieves greater execution efficiency than baseline methods. It also delivers higher folding accuracy in two tasks while maintaining competitive performance in the third. Despite being trained only on simulated rectangular cloth, the system generalizes well to real-world manipulation of nonrectangular garments such as T-shirts and shorts, requiring only minimal fine-tuning. The demonstration video is available at: https://www.youtube.com/@Jaui-g9j. Originality/value This study presents a practical dual-arm folding system featuring a novel perception architecture that yields both high accuracy and exceptional sample efficiency for sim-to-real transfer. By focusing on structural feature learning via multiresolution attention and direct action prediction, HrcbamFolding advances the state-of-the-art toward generalized and data-efficient robotic fabric manipulation, offering significant value for real-world automation.