ContactMPC: Towards Online Adaptive Control for Contact-Rich Dexterous Manipulation

Arjun Lakshmipathy, Nancy S. Pollard

Track: Paper Submission Track

Keywords: Control for Dexterous Manipulation, Model-Predictive Control, Contact Modeling

TL;DR: We present an online MPC framework for contact-rich dexterous manipulation by exploiting expected contact distributions and B-Spline control point exploration

Abstract: Model-based controllers are appealing due to their speed, transparency, and ability to adapt to new situations without strong reliance on extensive previously compiled data. However, online model-based control for dexterous manipulation remains challenging due to the need to navigate a large number of degrees of freedom through complex contact dynamics in real time. In this paper, we show that a simple sampling-based online model-predictive control (MPC) framework can be made to work even for complex multi-step tasks such as picking up a stapler from a table and reconfiguring it in-hand for use. We show that three elements are key to making this approach work in practice. First, a single reference motion is provided, which can come from a source other than the robot. In our case, we provide human motion capture data as reference for control of an Allegro robot hand. Second, contact area information from the reference motion is digested and made part of the loss function during sampling. Third, samples for exploration are taken in PCA basis directions derived from the reference motion. Ablation tests show that all three of these elements are needed to obtain successful results. In particular, sampling without contact information and sampling in the default configuration space of the robot result in uniformly low success rates for multiple tasks. We show examples of the Allegro hand performing several contact-rich motions in simulation, including twisting and pulling a doorknob, lifting an apple, lifting and pouring from a water bottle, and lifting and operating a stapler. We conclude with a discussion of implications for dexterous robotic interaction in the real world.

Submission Number: 185