An Autonomous Tensegrity Robot with Closed-loop Control and Real-time State Estimation

Tensegrity robots are lightweight, actuated structures composed of rigid struts and elastic tendons that can accommodate extreme deformations. Their ability to survive harsh impacts through deformations motivates tensegrity robots for navigating unstructured environments. However, tensegrity robots are limited by insufficient onboard sensing and a large sim2real gap. Here, we present an autonomous 3-bar tensegrity robot with elastic sensor tendons for closed-loop control and sim2real transfer. Highly deformable sensors for feedback control allow this robot to achieve high speeds (34 body lengths per minute), climb the steepest incline (28 degrees) of any tensegrity robot, navigate unstructured terrains (pebbles, grass, sand, etc.), and accurately reconstruct its shape and orientation (within 10% of the bar length) relative to Earth’s gravitational and magnetic fields. Finally, we demonstrate high-fidelity transfer of control policies from a differentiable physics-based simulation to the real robot in a real2sim2real pipeline where the physical parameters in the simulation are repeatedly updated based on data from the real robot.