Robust Planning for Dynamic Tensegrity Structures

This project develops high-performance motion planners for dynamic tensegrity robots. The planners consider the underlying dynamics of the structure so as to be robust to uncertainty due to soft contacts and compliance. Tensegrity is a good candidate to construct mobile robots with arbitrary forms that can adapt efficiently to highly unstructured and cluttered environments, such as uneven planetary surfaces. A tensegrity space exploration rover introduces small weight overhead relative to a science payload, it protects the key equipment and is able then to provide both surface mobility and manipulation capabilities. Controlling tensegrity robots is challenging due to high-dimensionality, non-linear dynamics and significant uncertainty. When a force is applied, non-linear and oscillatory responses arise due to interactions with the environment. These aspects make tensegrity structures difficult to control with traditional model-based approaches. These issues have limited existing solutions to quasi-static paths. This project creates new algorithms that overcome these challenges by taking advantage of progress in planning with dynamics and belief-space planning, especially in the case of non-Gaussian noise and highly non-linear dynamics. The project evaluates these solutions in the context of a tensegrity rover that navigates over an uneven terrain. Deviations relative to the original objective: Given developments in the research field and after coordinating with our NASA research collaborators, we have extended the project’s objective so as to include additional elements related to reinforcement learning on top of our original objective of providing kinodynamic planning solutions. Furthermore, for the last component period of the award (Jan. 2019 – Sep. 2019) the focus is on closing the gap between the simulation of the tensegrity systems on the NTRT software and alternative simulators (such as MuJoCo and the Bullet engine) with the real system through a data-driven approach. This will simplify the process into the future for researchers to experiment with tensegrity rovers in simulation and achieve the transfer of the results to the real system. 
•Our work on a framework for learning a differentiable physics-based engine for cable-driven robots from experimental data matured during the reporting period and eventually resulted in a publication that appeared in the 2nd Conference on Learning for Dynamics and Control (L4DC). We built a software infrastructure for experimenting with trainable physical models of tensegrity robots and evaluating the success of learning data-driven models of soft elements, such as cables and springs. 
•An undergraduate student, Mr. Patrick Meng, built two physical prototype of tensegrity robots in the lab with the objective of evaluating the physical attachment of multiple, modular robots. The work resulted in a paper accepted at the ASCE Earth and Space Conference 2020, with the undergraduate student as the first author and in collaboration with colleagues Devin Balkcom from Dartmouth, Weifu Wang from SUNY Albany.
•Following up on the work from this award, the PI submitted a collaborative proposal to NSF with Rebecca Kramer-Bottiglio (Yale – fellow NASA ECF awardee in the area of tensegrity), Devin Balkcom  (Dartmouth),  Weifu Wang  (SUNY  Albany).  The  proposal  has  been  recommended for funding (~$1.2M split across the 4 institutions for a period of 3 years) starting in September 2020.