Dexterous in-hand manipulation is an essential skill of production and life. Nevertheless, the highly stiff and mutable features of contacts cause limitations to real-time contact discovery and inference, which degrades the performance of model-based methods. Inspired by recent advancements in contact-rich locomotion and manipulation, this paper proposes a novel model-based approach to control dexterous in-hand manipulation and overcome the current limitations. The proposed approach has the attractive feature, which allows the robot to robustly execute long-horizon in-hand manipulation without pre-defined contact sequences or separated planning procedures. Specifically, we design a contact-implicit model predictive controller at high-level to generate real-time contact plans, which are executed by the low-level tracking controller. Compared with other model-based methods, such a long-horizon feature enables replanning and robust execution of contact-rich motions to achieve large-displacement in-hand tasks more efficiently; Compared with existing learning-based methods, the proposed approach achieves the dexterity and also generalizes to different objects without any pre-training. Detailed simulations and ablation studies demonstrate the efficiency and effectiveness of our method. It runs at 20Hz on the 23-degree-of-freedom long-horizon in-hand object rotation task.

翻译：灵巧手内操作是生产生活中的一项关键技能。然而，接触的高刚性和易变性特征限制了实时接触发现与推理，导致基于模型的方法性能下降。受近期接触丰富型 locomotion 与操作领域研究进展的启发，本文提出一种新颖的基于模型的控制方法，以克服当前灵巧手内操作的局限性。该方法具备显著优势：无需预定义接触序列或分离式规划流程，即可使机器人鲁棒地执行长时域手内操作。具体而言，我们设计了一个高层接触隐含模型预测控制器，用于生成实时接触计划，并由低层跟踪控制器执行。相较于其他基于模型的方法，这种长时域特性能够实现接触丰富型运动的重新规划与鲁棒执行，从而更高效地完成大位移手内任务；相较于现有基于学习的方法，本方法在实现灵巧性的同时，无需预训练即可泛化至不同物体。详细的仿真与消融实验验证了该方法的效率与有效性。在23自由度长时域手内物体旋转任务中，该系统以20Hz频率稳定运行。