Contact-Implicit Model Predictive Control for Dexterous In-hand Manipulation: A Long-Horizon and Robust Approach

The paper is submitted to IROS2024.

The source code will be released after the publication of the paper.

You can access the pre-print version on arXiv!

Video

Abstract

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.

Contact

If you have any question, feel free to contact the authors: Yongpeng Jiang, jiangyp19@gmail.com .

Yongpeng Jiang’s Homepage is at director-of-g.github.io.