An Open-Source, Reproducible Tensegrity Robot that can Navigate Among Obstacles

TL;DR

This paper tackles autonomous navigation for a 3-bar tensegrity robot in obstacle-rich, unstructured environments. It introduces a complete, open-source pipeline combining a differentiable physics-based system identification of motion primitives, a pose-estimation feedback loop, and planning over a discrete motion-primitive graph using $A^*$ in the $SE(2)$ space. Contributions include an open-source hardware design, an open-source software stack for modeling, planning, and control, and demonstrated robustness to disturbances such as vertical drops, inclines, and granular terrain, including outdoor field tests and reproducibility across two laboratories. The platform provides a practical baseline for advancing navigation on compliant, shape-morphing robotic platforms and for broader adoption by the robotics community.

Abstract

Tensegrity robots, composed of rigid struts and elastic tendons, provide impact resistance, low mass, and adaptability to unstructured terrain. Their compliance and complex, coupled dynamics, however, present modeling and control challenges, hindering path planning and obstacle avoidance. This paper presents a complete, open-source, and reproducible system that enables navigation for a 3-bar tensegrity robot. The system comprises: (i) an inexpensive, open-source hardware design, and (ii) an integrated, open-source software stack for physics-based modeling, system identification, state estimation, path planning, and control. All hardware and software are publicly available at https://sites.google.com/view/tensegrity-navigation/. The proposed system tracks the robot's pose and executes collision-free paths to a specified goal among known obstacle locations. System robustness is demonstrated through experiments involving unmodeled environmental challenges, including a vertical drop, an incline, and granular media, culminating in an outdoor field demonstration. To validate reproducibility, experiments were conducted using robot instances at two different laboratories. This work provides the robotics community with a complete navigation system for a compliant, impact-resistant, and shape-morphing robot. This system is intended to serve as a springboard for advancing the navigation capabilities of other unconventional robotic platforms.

Figures (14)

• Figure 1: The open-source tensegrity robot moving around obstacles outdoors starts from the top left and autonomously navigates to the bottom right.
• Figure 2: Open-source design: Each strut has 2 motor enclosures; each housing a brushed DC motor and LiPo batteries. A custom motherboard uses commercial-off-the-shelf circuits mounted on headers, allowing for easy replacement. A power switch and charging port provide convenience. The tensegrity robot is formed by assembling 3 struts with strain sensors.
• Figure 3: (Left) Offline sysID: Real trajectories are generated by chaining human-engineered gaits and tracking the robot's pose via vision. Then, the start state $\mathbf{X}_0$ and controls $\mathbf{U}_{0:T-1}$ are autoregressively inputted into a differentiable simulator to predict corresponding trajectories. A scalar loss function between the ground truth and predicted trajectories is back-propagated to update system parameters. (Middle) Offline primitive modeling: Multiple human-engineered and in-simulation-found gaits are executed in the identified simulator. The changes in the pose of each primitive are recorded and stored in a motion primitive library. (Right) Online loop: As the tensegrity moves, the pose is estimated at the start and end of each primitive motion. Then a plan from the tensegrity's current pose to the goal is computed, and its first primitive action executed. This is repeated until the tensegrity reaches the goal.
• Figure 4: Example open-loop plan (dark blue) around obstacles (light blue circles) with unused expanded configurations (red).
• Figure 5: Online perception-plan-execution: Pose tracking runs at 7 Hz. The pose is passed to an A* planner. The first primitive in the new plan is executed. Re-planning occurs at the last step of each primitive as the robot returns to its rest state.