Dual-arm cloth unfolding via a two-stage perception and deep reinforcement learning approach

Autonomous manipulation of deformable objects such as cloth remains a fundamental challenge in robotics due to high-dimensional geometry, nonlinear dynamics, and frequent self-occlusions. This paper presents a two-stage perception-and-control framework for dual-arm cloth unfolding that integrates edge-based visual processing with deep reinforcement learning for grasp selection. In the first stage, edge detection from RGB images generates candidate grasp points along salient cloth edges. A DQN-based policy then selects the optimal grasp pair from this dynamically constructed action set to maximize subsequent cloth spreading. In the second stage, a fixed bi-manual motion primitive executes a coordinated pick-lift-stretch action to unfold the garment. Real-world experiments conducted across multiple fabrics demonstrate strong stage-level success rates and consistent improvements in global flatness, particularly as measured by plane-residual depth metrics. Results also reveal that topological unfolding and geometric smoothing are distinct outcomes with the latter requiring more sensitive global metrics than depth variance alone. Overall, the proposed method shows robust generalization across diverse cloth types and establishes a foundation for future learning-based control of continuous dual-arm unfolding motions.