Zero-order optimization with contact priors for locomotion

Victor Dhédin, Majid Khadiv

Keywords: Contact planning, Trajectory Optimization, Contact-Rich Robotics, Zero-order optimization, Legged Robots

TL;DR: This paper investigates the use of zero-order optimization for planning trajectories that realize predefined contact sequences with legged robots.

Abstract:

In this work, we introduce a contact-explicit trajectory optimization framework adapted for zero-order optimization methods. Our approach optimizes state–control trajectories for predefined full-horizon contact sequences (including both contact status and contact surface). We demonstrate the effectiveness of our method on multiple quadruped locomotion tasks. Even though zero-order methods do not require any contact information, we show that explicitly providing it significantly improves performance in challenging scenario, such as clearing a long gap with two lateral walls support.
This approach is inherently parallelizable, which opens the door to large-scale parallel data collection for contact-rich tasks, an important direction given recent advances in imitation learning.

Submission Number: 14