Nonnegative Least Squares-Based Multicontact Models for Explicit Model Predictive Control of Dexterous Manipulation

In dexterous manipulation, learning-based methods have achieved success while requiring extensive data and lacking interpretability, whereas model-based methods are interpretable but costly due to the nested optimization in model predictive control (MPC). Recently, based on simplified algorithms, explicit model methods have overcome some difficulties and achieved satisfactory performance. In this article, a simplified nonnegative least squares (NNLS)-based quadratic programming (QP) algorithm is utilized to obtain the approximation solution to the dual QP of the contact model, with a hyperparameterized solution to the NNLS formulation used. As a result, an explicit and interpretable ContactNNLS model is proposed, eliminating the need to solve the nested optimization. The ContactNNLS model delivers a more general form when compared with the existing explicit models. Besides, the ContactNNLS model is integrated into an MPC to approximate the time-stepping prediction, leading to a ContactNNLS-MPC algorithm for dexterous manipulation. The effectiveness of the ContactNNLS model and the ContactNNLS-MPC algorithm is verified in extensive simulated tasks. Moreover, the ContactNNLS-MPC is further applied to a real Allegro hand for on-palm reorientation tasks, demonstrating its practicality and feasibility.