Dynamic Programming using Tensor Approximation for Contact-rich Manipulation

Teng Xue, Suhan Narayana Shetty, Sylvain Calinon

Keywords: approximate dynamic programming, tensor train, contact-rich manipulation

Abstract: Contact-rich manipulation is a challenging task for robot planning and control, due to the varying contact modes involved and the resultant uncertainties that arise from the contact points. Approximate Dynamic programming (ADP) is a promising approach for such problems as it can handle hybrid systems. Most existing methods in ADP that use neural networks for function approximation and gradient-based optimization procedures for policy retrieval struggle to handle the hybrid nature of contacts. In this work, we present a gradient-free, low-rank tensor approximation approach using Tensor Train (TT) to approximate the value function. The associated numerical optimization techniques for functions in TT format further allow performing optimization over both continuous and discrete variables, hence allowing handling hybrid systems. We demonstrate the effectiveness of our approach on a non-prehensile manipulation task with hybrid states and actions in both simulation and the real world.

Confirmation: I am a student and I would like to be considered for the funding award.