Learning to Unfold Garment Effectively Into Oriented Direction

In the real world, unfolding the garment to a specific plane direction and vertical facing can make the downstream task, i.e. folding, more convenient and effective. Then, we propose a policy that strategically selects action between the dynamic fling and quasi-static pick & place to effectively unfold any arbitrarily configured garment into two specific orientations with maximum coverage. In this work, we define a novel factorized reward function comprising garment coverage and two orientations (plane direction and novel proposed vertical facing) to train the policy. Moreover, we employ two prior knowledge modules: the Value Attention Module and the Action Optimized Module. The former assigns higher value weights to the key points of the garment, while the latter optimizes the lifting height and the flinging speed. Experimentally, we demonstrate the performance against three baselines in simulation. Our approach achieves two specific orientation configurations, especially in the vertical facing, which is not addressed by other methods. Furthermore, compared to the SOTA, we achieved approximately 4.0% and 17.6% improvements in coverage and manipulation steps, respectively. Our method is also finetuned on a real dual-arm robot to narrow the gap between the real world and simulation. Finally, our method is applied to a whole folding task as the initial unfolding step and demonstrates its performance.