Title: Towards Contact Robust Locomotion: Planning and Estimation Methods for Robot Locomotion under Uncertainty in Contact Conditions
Dr. Ye Zhao (Advisor) – George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology
Dr. Frank Dellaert – School of Interactive Computing, Georgia Institute of Technology
Dr. Seth Hutchinson – School of Interactive Computing, Georgia Institute of Technology
Dr. Jun Ueda – George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology
Dr. Patricio Vela – School of Electrical and Computer Engineering, Georgia Institute of Technology
Abstract: As robots move from the laboratory into real world environments, accounting for model uncertainty will become increasingly important. Current optimization-based motion planning techniques can produce dynamic robot motions for a variety of tasks; however, these method require a precise environment model, which can be intractable to generate for all real-world scenarios. To account for uncertainty in contact modeling during motion planning, we have investigated two separate robust planning techniques. First, we converted the uncertain contact constraints into a robust cost, and demonstrated that the resulting optimization reduces sliding time under frictional uncertainty and increases foot clearance under contact distance uncertainty. Second, we explored converting the stochastic constraints into deterministic chance constraints, and we combined them with our robust cost. We showed that the chance constraints alone do not add robustness, but can mediate the tradeoff between following the robust and nominal trajectories at constant contact uncertainty.
Inspired by our work on contact-robust optimization, our proposed work focuses on developing an optimization-based approach to update contact model uncertainty from executed trajectory data. We will develop our approach by inverting the complementarity model of contact; we will estimate the errors in the contact model and estimate the forces given the executed trajectory, and combine the errors with the prior contact model using Gaussian process regression. Our approach will be evaluated first in a simple sliding block example and then in a robotic quadruped example.