An Imitation Learning-Based Approach for Enabling Plant-Like Tropic Abilities in a Redundant Soft Continuum Arm

• Published in: 2024 10th International Conference on Control, Automation and Robotics (ICCAR) pp. 101 - 108
• Main Authors: Nazeer, Muhammad Sunny, Ansari, Yasmin Tauqeer, Riviere, Mathieu, Meroz, Yasmine, Laschi, Cecilia, Falotico, Egidio
• Format: Conference Proceeding
• Language: English
• Published: IEEE 27.04.2024
• Subjects: Adaptive Control; Control; Imitation Learning (IL); Inverse Kine-matics; Kinematics; Manipulators; Plant Tropism Movements; Real-time Control; Real-time systems; Shape; Soft robotics; Soft Robots; Stochasticity; Tendon-driven Soft Robots; Training; Trajectory tracking
• ISSN: 2251-2454
• DOI: 10.1109/ICCAR61844.2024.10569321

Plants, despite their sessile nature, have evolved growth-driven movement strategies that make them highly adept at negotiating with a wide range of highly uncertain environmental conditions and terrain. Replicating these capabilities is promising to endow soft robot arms with a novel repertoire of kinematics motions, thereby, enabling their deployment to highly unstructured environments. From among various behaviors, this work takes inspiration from tropisms, which is a bending response plants employ to rapidly move towards a desired external stimulus. The idea is to emulate these motions in order to achieve a desired trajectory tracking in redundant soft robot arms, a known nontrivial problem. Interestingly, these motions mathematically operate as optimal trajectories, thereby, laying the foundation to formulate an imitation-learning based control paradigm. In particular, we use data from experiments on wheat coleoptile shoots which generate optimal curvature away from the direction of gravity even under harsh environmental conditions. We tested and validated the proposed controller on a 9 DoF modular soft continuum arm, both in simulation and hardware. We demonstrate that with only a few natural trajectories, we can learn an imitation policy that not only enables trajectory tracking over the original task, but also, generalizes to other trajectories with an accuracy of 1cm, on average.