Learning Contact-rich Abstractions using Tensor Factorization

Teng Xue, Amirreza Razmjoo, Suhan Shetty, Sylvain Calinon

Keywords: Abstraction Learning, Contact-rich Manipulation, Tensor Train

TL;DR: We propose contact-rich abstraction learning through tensor factorization for long-horizon planning and control.

Abstract: Contact-rich manipulation is non-trivial due to its under-actuated dynamics and hybrid contact modes, which lead to non-convex programs and require joint reasoning over both discrete and continuous variables. This presents significant challenges for current gradient-based and sampling-based methods, especially in long-horizon manipulation tasks with sparse rewards. Planning with effective abstractions is a promising approach to address these issues. However, identifying an effective abstraction of the geometric world that can facilitate motion-level planning and control remains an open question. In this work, we propose learning such abstractions through tensor factorization, which enables efficient planning and control in contact-rich scenarios. We validate the proposed method across three manipulation domains, encompassing both prehensile and non-prehensile primitives. The results demonstrate its ability to find the optimal solution over the full logic and geometric path. Real-robot experiments further showcase the effectiveness of our approach in handling contact uncertainty and external disturbances in the real world.

Submission Number: 11